Impact of Service Sector Loads on Renewable Resource Integration

Urban areas consist of a mix of households and services, such as offices, shops, schools, etc. Yet most urban energy models only consider household load profiles, omitting the service sector. Realistic assessment of the potential for renewable resource integration in cities requires models that include detailed demand and generation profiles. Detailed generation profiles are available for many resources. Detailed demand profiles, however, are currently only available for households and not for the service sector. This paper addresses this gap. The paper (1) proposes a novel approach to devise synthetic service sector demand profiles based on a combination of a large number of different data sources, and (2) uses these profiles to study the impact of the service sector on the potential for renewable resource integration in urban energy systems, using the Netherlands as a case study. The importance of the service sector is addressed in a broad range of solar and wind generation scenarios, and in specific time and weather conditions (in a single scenario). Results show that including the service sector leads to statistically significantly better estimations of the potential of renewable resource integration in urban areas. In specific time and weather conditions, including the service sector results in estimations that are up to 33% higher than if only households are considered. The results can be used by researchers to improve urban energy systems models, and by decision-makers and practitioners for grid planning, operation and management}.


Introduction
Urban areas are major energy consumers [1]. In the coming decades, urban areas will need to transition to renewable energy resources to satisfy their demand in a sustainable manner [2,3]. Urban demand broadly falls in three categories: industrial, residential and service sector. This paper considers only the latter two since industrial consumers are typically located on city edges and have case-specific requirements to transition to renewable supply [4,5]. Considerable work has already been done on characterising and modelling residential demand [6,7]. The service sector has received much less attention. Yet service sector loads have markedly different demand profiles as compared to households [8,9]. Disregarding their presence in urban areas is misleading when assessing the effects of renewable resource integration.
The importance of service sector consumers for power balancing in future urban grids is increasingly acknowledged in literature [10,11,12,13,14]. However, most modelling studies are based on residential loads only, with some notable exceptions, such as [15], who take some service sector loads into account. Studies on service sector demand primarily focus on subsector energy use [8,9,16,17,18] (see further). Overall, little literature is available on the effects of service sector loads on the integration of renewable resources. This paper addresses this hiatus by explicitly modelling service sector loads as an inherent part of urban demand. Different renewable resource integration metrics are compared for two load types: (1) mixed residential and service sector loads in realistic proportions, and (2) residential loads only. This is the first fundamental study of the impact of service sector loads on renewable resource integration.

Service Sector Demand
The service sector, also termed commercial, business or tertiary sector, comprises a highly heterogeneous group of power consumers. Sector definitions differ, but mostly include nonmanufacturing commercial activities and exclude agriculture and transportation [9,18,19]. This paper defines the service sector as the collection of non-manufacturing commercial and governmental activities, excluding agriculture, transportation, power sector, street lighting and waterworks.
The service sector power demand in developed countries currently accounts for one quarter to one third of the total national power demand, and is thus on par with residential demand [20,21,22]. In 2050, the demand shares of the service and the residential sectors are projected to increase to 40% each, at the expense of the industry demand, which will account for only 20% of the total national demand [21,23,24].
Although the service sector demand is increasingly important, its load characteristics are poorly understood. The sector's heterogeneity and its historically modest share in the total national power demand are seen as the main causes for the existing lack of knowledge [18,19]. Some efforts are undertaken to remedy the situation. Recent studies seek to characterise power consumption per subsector and end-use. These studies are based on cumulative annual values [8,9,16,17,18], and mostly aim to estimate the effects of energy efficiency [8,17] and demand response [16] programs. Studies using hourly data are rare, with [9] a noteworthy exception. Yet such detailed (e.g., hourly) data are needed to study the interrelations between variably demand and variable non-dispatchable renewable supply, and thus to prepare the power system for the transition to renewable resources.

Importance of Detailed Demand Characteristics
The current power system is built on the paradigm of dispatchable generation which follows a variable immutable load [25]. The main load characteristics considered, are cumulative annual load and maximal load peak [26,27,28]. Power systems with a high share of intermittent renewables require the consideration of more detailed load characteristics. In such systems supporting measures, such as demand response and storage, are needed to safeguard the balance between supply and demand. To design such supporting measures appropriately, it is key to understand the extent and timing of the imbalances between renewable generation and demand. These imbalances depend on (1) the type of load, (2) the time (e.g., time of the day, day of the week), and (3) the weather (which governs renewable generation). The latter two can only be understood if detailed (e.g., hourly) data are available. A novel time and weather dependency classification system is presented, showing the potential of detailed demand data. This paper relies on modelled hourly service sector demand data provided by the United States Department of Energy (DOE) [29] as sufficiently detailed measured open source demand data are unavailable.

Contributions
This paper reports on three experiments. First, a broad range of solar and wind resource penetration scenarios is explored. Second, for a single scenario of installed solar and wind capacity, this paper zooms in on the differences between different times of the day, days of the week, and weather conditions. Third, the impact of load flexibility and storage availability is assessed for the same scenario as used in the second experiment. The main contributions of this paper are: 1. Extensive data combination and analysis yielding a quantitative model for urban demand, consisting of a realistic mix of service sector and residential loads.
2. Development of a novel time and weather dependency classification system.
3. Quantitative analysis of a broad range of renewable resource penetration scenarios with respect to the importance of the service sector for renewable resource integration.
4. Quantitative analysis of different times of the day, days of the week and weather conditions with respect to the importance of the service sector for renewable resource integration.
5. Quantitative analysis of the impact of load flexibility and storage availability with respect to the importance of the service sector for renewable resource integration.
The remainder of this paper is structured as follows. Section 2 presents the theoretical rationale for explicit consideration of service sector loads. Section 3 outlines the methods used for data collection and profile calculation. Section 4 provides more details on the three modelling experiments. Section 5 presents the results of these experiments, which are further discussed in Section 6. A final conclusion is given in Section 7.

Rationale
Urban areas are typically a mix of residential and service sector loads. Residential load profiles are more readily available, making them an attractive proxy for urban areas as a whole. However, residential and service sector load profiles differ considerably. Fig. 1 illustrates the difference in load profiles between (1) residential loads only and (2) mixed residential and service sector loads. This paper hypothesises that realistic load profile differences are large enough to lead to significant differences in renewable resource integration metrics. In this section a theoretical intuition supporting this hypothesis is developed from two perspectives: mismatch between renewable generation and demand, and renewable energy utilisation. These form the basis for the metrics used in the remainder of this paper.

Load Type Comparison
Two load types are compared, (1) residential loads only, denoted by L r , and (2) mixed residential and service sector urban loads, denoted by L m . Let h(t) and s(t) represent respectively household and service sector load over time. Then, Note that L r is scaled by a factor φ to ensure that 8760 t=1 L r = 8760 t=1 L m for hourly steps of t. Figure 1: Comparison of load profiles of residential loads only (orange) and mixed residential and service sector loads (blue) on an average weekday. The shaded area represents the cumulative load difference. Blue shaded area shows energy consumption underestimation by residential loads only as compared to mixed urban loads. Orange shaded area represents the contrary case. Note that for the mixed load the proportions of residential and service sector are chosen to be representative for the Netherlands.

Mismatch
Generation and load must be in perfect balance for the proper operation of the power system. Let mismatch M be the difference between generation G and coinciding load L, thus M = G − L. For a given generation G, the difference in mismatch between calculations considering residential and mixed load equals: From Eq. 3 follows that the difference in mismatch ∆M only depends on the loads h(t) and s(t), not on the generation G. Fig. 1 shows that ∆M > 0 during the day and ∆M < 0 in the evening. This suggests that in urban areas with a mixed demand, power imbalance calculations based on residential load only will lead to underestimations of the mismatch between supply and demand during the day and overestimations of this mismatch during the evening. Numerical values and statistical significance of these errors are shown in the following experimental sections.

Renewable Energy Utilisation
A similar analysis can be carried out for renewable energy utilisation, denoted by R. Assuming no storage or load flexibility, R is computed as: Given a generation G, the difference in renewable energy utilisation between residential loads only and mixed loads equals: Expanding Eq. 5 yields ∆R as a function of both renewable generation and load. As both are time and weather dependent, assessment of renewable energy utilisation requires an analysis of time and weather interactions, in addition to correct load profile estimation. In this paper a model is presented which allows to study the influence of load type, both on an annual basis, and in specific time and weather conditions. This model is described in the next section.

Simulation Model
This paper uses an extensive data collection and simulation approach to assess the impact of service sector loads on renewable resource integration. Measured detailed service sector load data are scarce, therefore a large number of data sources is combined to create a detailed realistic data series. To obtain numerical results, the Netherlands is chosen as a case study. The influence of load type, and of time and weather on renewable resource integration metrics is studied using a novel simulation model (developed in Matlab [30]). Load type effects are assessed by comparing two load cases: residential loads only, and mixed residential and service sector loads. Time and weather effects are studied using a novel time interval classification system. The approach is conceptually shown in Fig. 2. It consists of three steps: (1) data collection, (2) profile modelling, and (3) simulation experiments. The first two steps, the core of the simulation model, are outlined in the next section. The experiments carried out are described in Section 4.

Data Collection and Profile Modelling
Synthetic load and generation profiles are calculated based on a large number of data sources, detailed below. To ensure spatial and temporal consistency, all calculations are done for the same area (Amsterdam, the Netherlands) and the same period (2014), taking into account official Dutch holidays and daylight saving times. The network is assumed to be a "copper plate". All resulting profiles have an hourly granularity.

Load Types
Two types of loads are defined: residential load (comprised of household loads only) and mixed load (a mix of household and service sector loads). For both load types, household load is represented by a single average Dutch household profile. For the mixed load type, service sector load is calculated as a weighted sum of thirteen subsector load profiles (described below in more detail). To ensure that comparison between the two load types is fair, an equal annual cumulative consumption (718 GWh/year) is used for both load types (see also 1). To achieve this, the residential load is weighted by a factor φ (for the Netherlands, φ = 2.05, see also Eq. 1).
3.1.1.1. Household Load Profile. Household demand data are obtained from [31]. The average yearly household consumption is assumed to be 3500 kWh [32]. The selected profile describes the average Dutch residential load. The use of this single profile is deemed representative at the scale used in the simulations in this paper (100 000 or 205049 households, depending on load type).

Service Sector Load Profiles.
Sufficiently detailed open source service sector consumption data are only available from the United States Department of Energy (DOE) [29]. These data are used in this paper. The DOE provides data for 16 types of commercial consumers. Of these, 13 types are considered in this paper (as listed in Table 1). The remaining three (Midrise Apartment, Outpatient Health Care and Strip Mall) are not used due to their poor matching within the Dutch context.
The available consumer types are assumed to represent subsectors of the service sector. To create a single realistic service sector profile for the Netherlands, the profiles of the 13 consumer types are combined. Each consumer type reference building, for instance a hospital, is scaled to represent the whole subsector, in this case healthcare. These scaling factors are expressed as the number of U.S. reference buildings necessary to realistically account for a given Dutch subsector load (e.g., 4 hospitals to account for healthcare for a small city of 100 000 households). Due to a lack of consistent data, the scaling factors are based on different data, such as floor area, number of employees, number of students, etc. An overview of the service sector consumer types and corresponding data is given in Table 1.
The service sector profiles themselves are obtained using the DOE EnergyPlus modelling software [33]. This software builds demand profiles based on the building age, climate data, and the building location. As the simulations assume a future situation, new construction (post-2004) standard is used. To create profiles representative for the Netherlands, Amsterdam climate data are used [34]. Finally, the location match in terms of climate zone is based on both the ASHRAE climate classification [35] and the available U.S. locations for the reference models, yielding Seattle as the closest match for Amsterdam.

Generation Profiles
Solar and wind power generation are modelled using weather data from the Royal Netherlands Meteorological Institute (KNMI) [51]. Table 1: Service sector consumer types considered in this paper. A single representative service sector profile is constructed through the combination of the profiles of the different consumer types. Each consumer type reference building is scaled to represent a subsector. The number of reference buildings necessary to model the entire subsector is estimated based on different data sources, as consistent data are not available. The number of reference buildings is based on a small city with 100 000 households. The roof area is the area available per building for solar panels. References for subsector scaling factors are listed in the last column.

Service Sector
Number  [52]. The technical specifications are based on Solarex MSX-60 panels [53]. This paper assumes that solar panels can be placed on roofs of residential and service sector buildings. The roof area is used as a constraint to the number of solar panels which can be used. For the service sector, the maximal available roof area is calculated as the ratio between the total floor area and the number of storeys [29]. For households, an average roof area of 33 m 2 is used [54]. All roofs are assumed to be flat and allowing optimal positioning of solar panels.

Experiments Description
Three simulation experiments are carried out to study the impact of service sector loads on renewable resource integration. This impact is quantified using four metrics. The next paragraph outlines the metrics used, the subsequent paragraphs provide details for the three experiments.

Metrics
The following four renewable resource integration metrics are used in this paper: Positive Mismatch. Positive mismatch accounts for generation excess. It is calculated as the difference between generation and load when generation exceeds load.
Negative Mismatch. Negative mismatch accounts for generation shortage. It is calculated as the difference between generation and load when load exceeds generation.
Renewable Energy Utilisation. Renewable energy utilisation is the amount of renewable energy which can be used by the coinciding load. It is assumed that whenever renewable energy is available, it is utilised first. Only if no renewable energy is available, other (non-modelled) sources are used.
Self-Consumption. Self-consumption is the ratio of renewable energy utilised by the coinciding loads to the total renewable energy generated.

Experiment 1: Renewable Resource Penetration Scenarios
Renewable resources considered in this paper are solar photovoltaic (PV) panels and wind turbines. For both solar PV and wind turbines, the installed generation capacity is varied between 0 MW and 530 MW with steps of 53 MW (121 scenarios in total). In case of residential load, 530 MW represents 300% of peak load (177 MW). In case of mixed load, 530 MW is 368% of peak load (144 MW), as mixed load has a flatter profile (see Fig. 1). The considered capacities are comparable to [21], where renewable resource capacity of up to 341% of peak load is considered for 2050.
In each scenario, the corresponding generation profile is calculated. This generation profile is combined with, on one hand, the demand profile of residential loads only, and, on the other hand, with the demand profile of mixed loads. For each scenario and for each load type, a year-long hourly simulation is run. From the results, annual metrics are calculated and reported.

Experiment 2: Time and Weather Dependency
For a single scenario of solar PV and wind turbine penetration, this paper zooms in on the role time and weather conditions play on the importance to consider the service sector in systems with high renewable resource penetration. A novel time and weather dependency classification system is introduced to study the impact of different days of the week, times of the day, and weather conditions. The single scenario of solar PV and wind turbine penetration is obtained as a result of an area-constrained optimisation.

Time and Weather Dependency Classification System
In a power system with a high penetration of renewables, not only load variations, which mainly depend on the time of the day and day of the week determine the system state, but also weather variations which govern renewable generation. To account for the future system dependency on both time and weather, this paper proposes a novel time and weather classification system. In this system, each hour of the year is classified according to four parameters: (1) day of the week, (2) time of day, (3) solar irradiance and (4) wind speed. Two categories are distinguished for the day of the week: weekday and weekend. Three categories are distinguished for the time of the day: night (00:00 -08:00), day (08:00 -16:00) and evening (16:00 -00:00). Five categories are distinguished for both solar power generation and wind power generation. In both cases, the categories are based on quantiles. In total, 150 time and weather dependent categories are defined. Their frequency of occurrence is summarised in Table 2.

Area-Constrained Renewable Mix Optimisation
The single renewable resource penetration scenario is based on an area-constrained optimisation. The optimisation problem is formulated as a constrained multi-objective non-linear problem with design variables x the number of solar PV panels and wind turbines: where: p pos : weighting factor of positive mismatch (p pos > 0, in this paper p pos = 1) p neg : weighting factor of negative mismatch (p neg > 0, in this paper p neg = 1) p ren : weighting factor of renewable energy utilisation (p ren < 0, in this paper p ren = −5) A roof : roof area available α : factor accounting for additional area available (φ ≥ 1, in this paper φ = 3) Positive mismatch M + (x), negative mismatch M − (x) and renewable energy utilisation R(x) are all function of the decision variables x P V and x turbine through their dependency on generation G. Generation G is calculated as: G = x P V · g P V + x turbine · g turbine , with g P V and g turbine respectively the generation profiles of 1 m 2 PV and one 500 kW wind turbine. This paper assumes that the total area available for renewable power generation is three times the size of the cumulative roof area of all the buildings considered (i.e. α = 3). The roof area itself is only available for solar power generation, while at most twice the roof area is available for wind power generation (α − 1 in Eq. 6). The footprint of a wind turbine is assumed to be 0.345 km 2 /MW [57]. The optimisation problem is solved using the genetic algorithm in Matlab.

Experiment 3: Flexibility
In future power systems additional flexibility measures, such as demand response and storage, are expected to be implemented. The effects of these measures on renewable power integration metrics are assessed by assuming that part of the load is flexible and limited storage is available. The results are compared to those from experiment 2.

Demand Response
This paper assumes that a fixed percentage of the load can be shifted. Based on literature review it is estimated that on average about 20% of both household and service sector demand is shiftable for 2 hours, both to an earlier and to a later consumption time [9,10]. Within these constraints, it is assumed that loads are shifted as to mitigate the largest positive and negative mismatches first, i.e. to reduce residual peak supply and peak demand.

Storage
This paper assumes that some of the excess energy can be stored in batteries. The total storage is assumed to be 600 MWh (i.e. sufficiently large to supply the average demand for 7.3 hours). Grid-to-battery efficiency is assumed to be 80%, battery-to-grid efficiency 90%. When demand response and storage are combined, it is assumed that demand response capabilities are used first as they can mitigate shorter-term fluctuations, while storage is used after all demand response capabilities are exhausted.

Statistical Analysis
Metric differences between residential loads only and mixed loads are analysed for statistical significance using the two-sample t-test. Since multiple scenarios or categories are compared at once, the significance level is corrected using the Holm-Bonferroni correction to control the familywise error rate at 5%. For the first experiment, the correction is made for 121 comparisons. For second and third experiments, the correction is made for 150 comparisons.

Results
This section presents the results of the three experiments conducted. The overarching aim is to study what the influence of service sector loads is on renewable resource integration. The first experiment addresses this question over a broad range of renewable resource penetration scenarios. The second experiment zooms in on the differences between different days of the week, times of the day, and weather conditions for a single scenario. The third experiment addresses the influence of load flexibility and storage availability, further adhering the same assumptions as used in the second experiment. : Annual average differences between residential loads only and mixed residential and service sector loads. The XY plane represents scenarios of wind and solar penetration (expressed as percentage of peak load assuming mixed loads). Statistically significant differences are shown as red areas. Simulations for residential loads only assume 205 049 households. Simulations for mixed loads assume 100 000 households and the corresponding number of service sector consumers as shown in Table 1. Note that the cumulative annual demand in both load types is equal (718 GWh/year). Figure 3 shows annual average differences between residential loads only and mixed residential and service sector loads for four renewable resource integration metrics across a broad range of renewable resource penetration scenarios. Scenarios with solar and wind generation capacity of up to 530 MW are considered, i.e. 300% of peak load for the residential loads only and 368% of peak load for the mixed loads (mixed loads have a flatter profile, see Figure 1). Figures 3a and 3b show respectively the annual average positive and negative mismatch differences between residential loads only and mixed loads.

Mismatch
Positive mismatch represents renewable generation excess, i.e. renewable energy which cannot be used by the local loads. Positive mismatch differences indicate to what extent renewable generation excess is larger for the case of residential loads only, as compared to mixed loads. The larger and more positive the differences are, the more excess for residential loads. When solar and wind penetration equals zero, the positive mismatch difference is zero, since no renewable power is generated. For all other penetration scenarios, differences increase with increasing solar penetration, while the variation as a function of wind is limited. Statistically significant positive mismatch differences are found for solar penetration levels above 74% of peak load and for wind penetration scenarios below 74% of peak load. It should be noted that these cut-off values are based on the scenario step granularity of 37% of peak load.
Negative mismatch represents generation shortage, i.e. additional energy to be supplied by non-renewable resources. Negative mismatch differences indicate to what extent generation shortage is larger for the case of residential loads only, as compared to mixed loads (more negative differences indicate larger shortage for residential loads only). Negative mismatch difference is zero when solar and wind penetration equal zero as no renewable generation is available for either load type. Negative mismatch is larger for residential loads only than for mixed loads, leading to negative mismatch differences below zero across all remaining scenarios. Negative mismatch differences are statistically significant for scenarios with solar penetration above 147% of peak load (at wind capacity above 37% of peak load). Figure 3c shows renewable energy utilisation differences between residential loads only and mixed loads. Renewable energy utilisation is the amount of renewable energy that can be used by the coinciding demand. Renewable energy utilisation differences indicate to what extent less renewable energy is used directly by the residential loads only than by the mixed loads (more negative differences indicate less utilisation by residential loads only). Renewable energy utilisation differences follow the same pattern as negative mismatch differences. For all scenarios, renewable energy utilisation is higher for the mixed loads than for the residential loads only, thus the renewable energy utilisation difference is negative. Statistically significant differences are found for scenarios with solar capacity at or exceeding 147% of peak load, at any wind penetration. Figure 3d shows self-consumption differences between residential loads only and mixed loads. Self-consumption is the ratio of renewable energy utilised by the coinciding demand relative to the total renewable energy generated. Self-consumption differences indicate to what extent less of the generated renewable energy can be used by the residential loads only, as compared to mixed loads (more negative differences indicate less self-consuption by residential loads only). Self-consumption is highest at very low penetration of renewable technologies and undefined for zero penetration. If only a small amount of renewable power is generated, there is a high probability that coinciding load will be sufficiently high to use it entirely, irrespective of the load profile. Self-consumption differences have a similar pattern as renewable energy utilisation differences, although differences at low wind penetration scenarios are more pronounced. Statistically significant differences are found for low solar capacity scenarios above 74% of peak load and for wind penetration of at most 147% of peak load.

Summary
Differences in renewable resource integration metrics between residential loads only and mixed loads are found across a broad range of scenarios. Statistical significance differs between metrics, yet is found in all scenarios except low solar, very high wind. All metrics are dependent on the presence of renewable generation and thus tend to zero at low solar and low wind penetration scenarios, leading to non-significant results. Overall, this experiment shows a significant difference in annual average metrics between residential loads only and mixed loads. It demonstrates that (1) metrics differ significantly between residential and mixed areas with the same annual cumulative load, and that (2) in mixed load urban areas assessment of renewable power integration metrics based on residential load only leads to significant errors.
The relative magnitude of these average annual differences is relatively small, up to approximately 5% of the total annual load. However, the differences between the metrics for residential loads only and mixed loads vary throughout the year, depending on both time and weather conditions. These variations are assessed in the next experiment.

Experiment 2: Time and Weather Dependency
A power system with a high penetration of renewables is highly dependent on both time and weather. To study this dependency, all hours of the reference year (2014) are classified using the time and weather classification system introduced in this paper. Each category has four parameters: day of the week, time of the day, solar generation, and wind generation. In total, 150 time and weather dependent categories are analysed. An example of a category is: all weekday night (0:00 -8:00) hours with solar generation between 0% and 3% of the installed capacity and wind generation between 0% and 5% of the installed capacity (this category is indicated on The results shown are obtained assuming an optimal renewable mix for the mixed loads: 399 MW solar PV and 30 MW wind turbines (i.e. total renewable capacity amounting to 243% of peak load in case of residential loads only and 298% of peak load in case of mixed loads). Figure 4 shows mismatch dependency on time and weather and compares residential loads only and mixed loads. Positive mismatch indicates renewable generation excess. Negative mismatch indicates renewable generation shortage. Mismatches closer to zero are better.

Mismatch
During weekdays and on weekend nights (Figure 4a-d), the mismatch is more positive for the residential loads only than for the mixed loads. In the weekends during the day and in the evening (Figure 4e-f), the mismatch is more positive for the mixed loads, although the differences are relatively small compared to the weekday categories. The largest differences occur on sunny weekdays at daytime (Figure 4b-c), and amount to up to 21% of the peak load (for mixed loads).
On an annual basis, 62% of positive mismatches occurs during weekdays at daytime when solar generation exceeds 40% of installed capacity, which corresponds to 7% of the time. Most negative mismatches (46%) occur during weekdays in the evening with solar generation below 3% of installed capacity, which corresponds to 20% of the time.
Statistical significance is not shown in the graph, yet is calculated as described in Section 3. Significant differences between mismatch results for the two load types are found for all data points on weekdays during the day (Figure 4b), as well as weekday and weekend evenings (Figure 4c and f) for low solar (generation below 3% of installed capacity). In other periods, statistically significant differences occur for some categories. The disparity in statistical significance between periods can be attributed to two reasons: the number of data points and the relative difference between residential loads only and mixed loads for a given period. First, since weather patterns are not dependent on the day of the week, weekdays have on average 2.5 times more data points per weather category than weekends. Second, during weekends and during night periods, the difference between residential loads only and mixed loads is smaller than during other periods since most service sector activities are shut down. Each category thus has four parameters: day of the week, time of the day, solar generation, and wind generation. For example, the category indicated by the red arrow represents all weekday night (0:00 -8:00) hours of 2014 with solar generation between 0% and 3% of installed capacity and wind generation between 0% and 5% of installed capacity. The average mismatch in these hours is -41 MW for both load types. The values on the x-and y-axes are quantiles. Note that some high solar categories are missing because they do not occur in the modelled reference year. The results shown assume 399 MW solar PV and 30 MW wind turbines as installed capacity. Figure 5 shows renewable energy utilisation dependency on time and weather and compares residential loads only and mixed loads. Renewable energy utilisation is the amount of generated renewable energy which can be used by the coinciding loads directly. Higher renewable energy utilisation is better.

Renewable Energy Utilisation
At low solar and wind power generation levels, the differences between residential loads only and mixed loads are small, both on weekdays and in weekends and during all times of the day. Differences increase as solar generation increases. Wind generation has limited effects as it represents only a small portion of the total renewable generation due to area constraints (see optimisation problem definition in Section 3). At higher solar generation levels, renewable energy utilisation is higher for the mixed loads than for the residential loads only. The service sector consumption profile coincides better with the solar power generation profile as both peak during the day. During the weekend at day and evening times (Figure 5e-f), the renewable energy utilisation at high solar irradiance levels is higher for residential loads only since many service sector loads do not operate during the weekend. The largest differences occur on sunny weekdays at daytime and in the evening (Figure 5b-c), and amount to up to 13% of the peak load. Statistically significant differences are found during weekdays at high solar generation levels for all periods (Figure 5a-c). Most renewable energy utilisation (26%) occurs during weekdays at daytime with high solar generation levels (above 40% of installed capacity), these categories correspond to 7% of the time. Further, 7% of the renewable energy is consumed during night and evening periods with lowest sun and highest wind (occurring 10% of the time). Figure 6 shows self-consumption dependency on time and weather and compares residential loads only and mixed loads. Self-consumption is the amount of renewable energy utilised relative to the amount generated.

Self-Consumption
Similarly to the mismatch and renewable energy utilisation metrics, during weekdays and weekend nights the mixed loads performs better than the residential loads only (Figure 6a-d).
During weekend days and evenings the opposite is the case, although differences are again small (Figure 6e-f). Similarly to other metrics, the largest differences are found on sunny weekdays at daytime and in the evening (Figure 6b-c), mixed loads yield a self-consumption up to 11% higher than residential loads. Statistically significant differences occur for similar categories as for mismatch. At low solar generation levels and at all wind generation levels, the self-consumption is 100%, meaning that all renewable power generated can be used by the modelled loads. As solar generation increases, selfconsumption decreases. During weekdays the differences between the two load types are biggest (Figure 6b). In these periods the self-consumption decreases faster for the residential loads only than for the mixed loads. This result illustrates that modelling only households underestimates the self-consumption of realistic mixed urban areas.

Result Dependency on Load Assumptions in the Optimisation Step
The results presented above rely on a renewable resource generation mix obtained by solving an optimisation problem assuming mixed loads. In this paper, the optimisation is constrained by area (see Section 4.3.2). This is the binding constraint for the number of wind turbines, regardless of the load type assumed. However, the optimal solar generation capacity changes with the load type. It is 15% lower if residential loads only instead of mixed loads are assumed. The general trends for time and weather dependency as shown in Figures 4 -6 remain similar if residential loads only instead of mixed loads are assumed. However, overall mismatches become more negative, renewable energy utilisation decreases and self-consumption increases.

Summary
Renewable power integration metrics vary as a function of both time and weather. The results shown rely on the proposed time and weather classification system. Pronounced solar generation dependency is found for all metrics due to the high share of solar PV in the generation mix. Relative metric performance of residential loads only and mixed loads differs per period. Overall, on weekdays (subplots a-c on Figures 4, 5 and 6) mixed loads leads to lower mismatches and higher renewable energy utilisation, on the weekends (subplots d-f on Figures 4, 5 and 6) the contrary is the case. This difference can be attributed to service sector operation hours. Statistically significant differences between residential loads only and mixed loads are primarily found on weekdays due to a larger number of datapoints per category and a larger difference between the two load type profiles. Overall, these results show complex interdependencies between time, weather and load type.

Experiment 3: Flexibility
New technologies, such as storage (e.g., electrical vehicles) and flexible loads capable of participating in demand response (e.g., smart appliances) are expected to penetrate the power system, alongside with renewable power generation. Their integration in the grid is expected to influence the utilisation of renewable energy. Here, the influence is explored as a function of time and weather dependency. The approach used in this experiment is identical to experiment 2, the results should be compared with those obtained in the previous experiment.
For this experiment it is assumed that 20% of both household and service sector demand is shiftable for 2 hours, both to an earlier and to a later consumption time. The total storage is assumed to be 600 MWh (i.e. sufficiently large to supply the average demand for 7.3 hours). Figure 7 shows the mismatch between supply and demand after demand response and storage, comparing residential loads only and mixed loads. This figure can be compared to Figure 4. Similarly to the case with no demand response and storage, on weekdays and weekend nights (Figure 7a-d) the mismatches remain more positive for the residential loads only. The contrary remains the case during other periods (Figure 7e-f).

Mismatch
Overall, the absolute value of mismatches decreases as compared to the case without storage and demand response. The largest decrease in positive mismatch is observed during high sun weekday and weekend nights (by up to 133 MW or 100%) and high sun weekday daytimes (by up to 108 MW or 65%). Negative mismatches decrease only slightly during weekday and weekend daytimes, indicating that the available flexibility does not suffice to supply the demand during daytimes with little renewable generation. Notably, negative mismatches do decrease on weekday and especially on weekend evenings (by respectively 36% and 60%). The largest remaining differences in mismatch are found on sunny weekday evenings (Figure 7c), and amount to up to 24% of peak load.
The number of statistically significant categories decreases, although statistically significant differences remain across all periods. Most negative mismatches (43%) occur during weekday evenings with low solar power generation (corresponding to 20% of the time). Due to demand flexibility, positive mismatches (i.e. generation excess) decrease considerably (from 164 GWh/year to 36 GWh/year). The remaining mismatches occur primarily during high solar weekdays and weekends at daytime (accounting for respectively 56% and 33% of the annual positive mismatch). Figure 8 shows renewable energy utilisation after demand response and storage, comparing residential loads only and mixed loads. This figure can be compared to Figure 5. A first difference between the two figures is the shape of the surfaces. With flexible load and storage, renewable energy utilisation does not level off when renewable generation is high. During weekday daytimes ( Figure 5 and 8b) with high solar power generation, renewable energy utilisation doubles. In similar categories during weekend days ( Figure 5 and 8e) renewable energy utilisation increases with 150%. Renewable energy utilisation remains low during low sun nights and daytimes, but increases fifteen fold in low sun, low wind evenings (from around 3 MW to 44 MW). A second difference between Figures 5 and 8 is the increased similarity between the two load type surfaces due to demand flexibility. Yet, some differences remain. The largest remaining differences in renewable energy utilisation are found on sunny weekday evenings (Figure 8c), and amount to up to 24% of peak load.

Renewable Energy Utilisation
Statistical differences between the two load types disappear in all but a few categories (high sun weekday daytimes and low sun evenings). Most renewable power is utilised during high sun weekdays at daytime (24%). Overall, 45% of all renewable power is used on weekdays between 08:00 and 16:00. The availability of local storage enables delayed renewable power consumption, for instance, 10% of all renewable power is utilised on weekday evenings during low sun periods.  Figure 9 shows self-consumption after demand response and storage, comparing residential loads only and mixed loads. This figure can be compared to Figure 6. With flexible load and storage, self-consumption increases from a minimum of 33% to a minimum of 54%. Renewable energy utilisation of 100% is achieved in more categories than without system flexibility. During high sun weekday and weekend daytimes, renewable energy utilisation is above 75% (Figures 9b  and 9e), compared to 33% without system flexibility (Figures 6b and 9e). Least self-consumption improvements are on weekday and weekend evenings (improvements of about 10%). The largest remaining differences in self-consumption are found on sunny weekday evenings (Figure 9c), selfconsumption is up to 20% higher for mixed loads than for residential loads only. Since overall the differences between load types diminish, the number of statistically significant differences decreases.

Summary
Demand flexibility and storage improve the three metrics considered: mismatches decrease and both renewable energy utilisation and self-consumption increase. Since generation capacity does not change, the metric trends remain primarily dependent on solar power generation levels. The differences between the two load types diminish due to demand flexibility. On an annual basis, self-consumption increases from 91% to 98%, corresponding to a rise from 324 GWh/year to 533 GWh/year. Negative mismatch (i.e. energy needed from other sources) improves from 394 GWh/year to 298 GWh/year. Positive mismatch (i.e. generation excess) decreases from 164 GWh/year to 18 GWh/year. Due to system flexibility, with the same renewable generation capacity, 74% instead of 45% of demand can be supplied by local renewable resources.

Discussion
In the coming decades, power systems are expected to transition to renewable generation. The European Union for instance envisions that up to 97% of its power will be generated by renewables by 2050 [3]. As urban areas are major power consumers [1], their power grid will need to undergo considerable transformations. A thorough understanding of system characteristics, for both demand and generation, is needed to facilitate such transition. This paper focuses on the demand side, in particular on the role service sector loads play in renewable resource integration in urban areas.
The service sector has thus far been mostly omitted in power system research. This paper is the first fundamental study of the impact of service sector loads on renewable resource integration. It compares renewable resource integration metrics for two types of loads: residential loads only and mixed residential and service sector loads.
Currently service sector demand data are not readily available. This paper proposes demand profile models based on available data. This section first discusses the modelling choices and validation. Next, the influence of service sector loads on renewable resource integration is discussed. Finally, the time and weather classification system, introduced in this paper to study time and weather dependency of renewable resource integration metrics, is evaluated.

Service Sector Load Profile Modelling and Validation
This paper relies on modelled service sector load profiles, since sufficiently detailed measured open service sector demand data are not available. The Netherlands is chosen as case study. The modelling choices and model validation are discussed next, followed by an assessment of possible generalisation of the obtained results to other countries.

The Netherlands as a Case Study
In the Netherlands, only cumulative average annual service sector electricity consumption data are openly available ( [32,58,59]). Therefore this study relies on U.S. DOE commercial reference building models. These models are used to generate realistic Dutch service sector profiles. The share of each subsector in the final mix is estimated using several cross-checked sources (see references in Table 1). Further, the best climate match possible between the U.S. models and the Netherlands is assured to ensure comparable heating and cooling requirements (see Section 3).
It is nevertheless difficult to estimate whether specific U.S. assumptions underlying the DOE reference building profiles cause deviations from real Dutch profiles. Perez-Lombard et al. [60] compared office energy end-use between U.S., Spain and UK. End-use differences exist between the three countries. The differences between U.S. and the two European countries are however not larger than between the two European countries themselves. This suggests that using U.S. data for the Netherlands does not lead to larger errors than using data from another European country. Although undesirable, the practice of using data from other countries is currently common due to limited service sector data availability [8].
Further comparison of the values obtained in this paper with those found in the literature is not straightforward. Data available from literature are cumulative annual values (e.g., [8,17]) or average daily profiles (e.g., [9]). Comparison with these data is hard due to both their lack of detail and the inconsistencies in service sector definitions, an issue also raised by others [8,18,60]. Therefore, obtained results are compared with cumulative annual Dutch service sector load data. The modelled service sector loads account for 70% of the cumulative Dutch service sector power consumption [32,58,59]. The remainder can be attributed to unaccounted for subsectors, inaccuracies in subsector share estimations and load profile deviations.
Given the lack of service sector data, model and result validation now relies on source combination and cross-referencing. To improve service sector modelling and model validation, three issues in this research field need to be addressed: (1) inconsistent service sector definitions, (2) lack of open data in general, and (3) lack of detailed (e.g., hourly) load profiles in particular.

Generalisation
The numerical results in this paper are based on the service sector consumer mix in the Netherlands (see Table 1). The composition of the service sector and its share in the total national demand differ between countries [9,17,18]. Yet the shape of the service sector demand profile, with a peak during the day, is similar across developed countries [8,29,60,61] and differs from the shape of household demand profiles, which typically peak in the evening [31,62]. The results obtained based on the Dutch service sector can most likely be generalised for other developed countries.
Based on the results presented in this paper, can be expected that the more important solar generation is in a country's renewable resource mix, the bigger the impact of service sector loads is. Solar power generation peaks during the day, matching better with the service sector demand peak than with the household demand peak.
To numerically validate the generalisation expectations, the presented model needs to be adapted for other countries and the simulations need to be repeated.

Impact of the Service Sector on Renewable Resource Integration
Despite the limitations posed by data availability, this paper is able to show the impact of service sector loads on renewable resource integration metrics. It thus provides a proof-ofconcept for realistic load modelling in urban areas, allowing an improved and more accurate assessment of renewable resource integration in urban power systems. The purpose is two-fold: (1) quantifying the errors made by estimating metrics for a mixed load area based solely on more readily available residential load profiles, and (2) assessing renewable resource integration metrics in areas with similar cumulative annual load, but with different load profiles differences arising from differences in load types.
The influence of load type is analysed by comparing residential loads only and mixed residential and service sector loads. Since the future generation mix is uncertain [63], first the influence of renewable energy penetration scenarios is addressed. Next, for a single generation mix scenario, the influence of time and weather is studied. Finally, the impact of load flexibility and storage is explored.

Experiment 1: Renewable Resource Penetration Scenarios
The impact of service sector loads on renewable resource integration is studied over a broad range of renewable resource penetration scenarios. Statistically significant differences between renewable resource integration metrics for residential loads only and mixed loads are found in all renewable resource penetration scenarios, except in those with high installed wind turbine capacity and low installed solar PV capacity, and those with few renewable resources (left upper side on each subplot in Figure 3). In most cases, mixed loads lead to less renewable energy excess, less energy requirements from other non-renewable resources, and thus to a higher renewable energy utilisation and higher self-consumption. Thus, given a similar annual load, mixed urban areas perform better in terms of renewable resource integration than what is expected based on residential loads only.
This paper assumes renewable resource penetration scenarios of up to 368% of peak load. This assumption is high, yet it is comparable to renewable generation capacities in other studies (e.g.up to 340% of peak load in [21]). Currently the Netherlands produces ten times more renewable power from wind (5300 GWh/year) than from solar (504 GWh/year) [64]. However, such relative proportions are unlikely at very high renewable resource penetrations. Wind turbines with an installed capacity of 368% of peak load of mixed residential and service sector in the entire Netherlands would cover approximately 30% of the land area. On the other hand, solar PV with an installed capacity of 368% of mixed loads for the entire Netherlands would cover only 1% of the land area. Thus, in the Netherlands, very high wind and low solar penetration scenarios are unlikely due to area constraints.
These considerations show that within the realistic range of renewable resource penetration scenarios in the Netherlands, service sector loads significantly influence renewable resource integration metrics. They should be taken into account when assessing the impact of renewable resource integration in urban power systems.

Experiment 2: Time and Weather Dependency
A renewable power system is highly dependent both on time and weather. Time governs diurnal, weekly and seasonal patterns in demand, and diurnal and seasonal patterns in solar generation patterns. Weather governs both solar and wind power generation, as well as some portion of the demand. The dependencies of renewable resource integration metrics on the interplay between time and weather are analysed using the time and weather dependency classification system introduced in this paper. This system itself is discussed further, here the focus is on the comparison between the metric trends obtained for residential loads only and for the mixed loads.
An initial inspection of the results presented in Figures 4 -6 shows that general trends are similar between residential loads only and mixed loads. All metrics depend on time and primarily on solar generation level (since solar power accounts for 93% of the installed renewable capacity due to area constraints).
Further statistical analysis reveals significant differences between the two load types in a number of time and weather dependent categories, in particular on weekdays during daytime, and for some metrics during the evenings and nights (depending on solar generation levels). Metric differences of up to 21% of peak load are found. These results thus supports the hypothesis that load profile differences between residential loads only and mixed loads lead to significant differences in corresponding renewable resource integration metrics.

Experiment 3: Flexibility
In the future, loads are expected to become partly flexible. In this study this is taken into account by assuming that 20% of the loads are shiftable up to two hours. Furthermore, storage capacity of 600 MWh is assumed, which corresponds to, for instance, 2000 households having an electric car with a 30 kWh battery (that is 1-2% of the households, depending on the load type). Load flexibility considerably improves all metrics. Renewable power consumption increases by 65% and covers 74% of total demand (up from 45%). Only 2% of the renewable energy cannot be consumed by the modelled loads. Most statistical differences between residential loads only and mixed loads disappear due to increased system flexibility. However, it should be noted that the model assumes that both the households and the service sector participate in demand response. Yet, existing demand response programs primarily target household consumers [5,6,65]. In literature only a limited number of studies exist on service sector demand response [9,16] and very few, if any, programs are implemented. If service sector loads remain untargeted by demand response programs, a considerable share of load flexibility will remain unharnessed leading to lower renewable energy utilisation and higher mismatches between generation and demand.

Time and Weather Dependency Classification System
The power system dependency on the weather increases generation becomes increasingly renewable. Although load also partly depends on the weather, currently power system metrics are assessed mainly from a time perspective [66]. This paper proposes a novel time and weather dependency classification system which takes both time and weather into account.
The time and weather dependency classification system is flexible and can be readily applied to a wide range of dataseries. In this paper the system application is shown for a reference case with time intervals of one hour, and full-year data. Classification parameter choices yielded 150 time and weather dependent categories. Time interval, dataseries size and number of categories can be varied, depending on the data and modelling purposes. For instance, the time and weather dependency classification system can be used with statistical data from multiple years to identify critical combinations of time and weather, to plan and dispatch system operations accordingly.
In this proof-of-concept example such critical combinations are reported for the reference year 2014 (see Section 5). The ability to identify such critical values as a function of time and weather and to assess their likelihood of occurrence is key for designing supporting measures (such as storage, load flexibility and dispatchable generation) for power systems with a high share of renewable resources.

Conclusions and Future Work
This paper seeks to contribute to the understanding of the impact of the service sector, an inherent part of urban areas, on the transition of urban power systems to renewable generation. So far, the service sector has been largely disregarded in power systems research. This is the first fundamental study addressing the impact of the service sector on renewable resource integration. The main contributions of this paper are: 1. The extensive data collection, combination and analysis which enables a realistic simulation of mixed service sector and residential demand. Currently, measured service sector load profiles are scarce. Therefore this paper relies on modelled profiles provided by U.S. DOE, which are adapted for the Netherlands, the chosen case study region.
2. This paper proposes a novel time and weather dependency classification system which allows metric analysis taking both time and weather into account. The existing approach primarily considers metric dependency on time, while power systems with a high penetration of renewables depend on both time and weather.
3. The impact of the service sector is shown in a broad range of solar and wind resource penetration scenarios. Statistically significant differences in renewable resource integration metrics between residential loads only and mixed loads are reported in all realistic scenarios.
4. Metric differences between residential loads only and mixed loads are analysed using the novel time and weather classification system. Largest differences between both load types occur on sunny weekdays at daytimes and in the evenings (i.e. in more than 1200 hours per year). Differences of up to 21% of peak load are found.
5. Renewable resource integration metric improvements are shown when part of the load is flexible and some storage is available. Load flexibility and storage availability diminish metric differences between the two load types. The largest remaining differences, up to 24% of peak load, are found on sunny weekday evenings. The simulated improvements in renewable resource integration metrics can only fully be realised in practice in mixed urban areas if dedicated demand response programs for service sector consumers are developed.
This paper is a first step towards a better understanding of the importance of the service sector in the transition of urban areas to sustainable power generation. Considerable work remains to be done. Similar analyses need to be carried out both on a smaller scales, i.e. for specific urban neighbourhoods as load characteristics vary locally, as on a larger scales, i.e. for other countries, as both generation and load conditions vary between geographical regions.
Current lack of open, detailed, measured service sector demand data thwarts its inclusion in power systems research. Therefore, detailed open service sector data and consensus building on definitions are indispensable for future research.