Short-term power-forecasting method of distributed PV power system for consideration of its effects on load forecasting

: This study analyses the distributed photovoltaic (PV) power system effects on common short-term load forecasting, based on the proposed power penetration index, and load data recon ﬁ guration method, by using common short-term load-forecasting algorithm under conditions of the distributed PV power system and the power penetration index as 0, 5, 10, 15, 20, 24.46, 30, 35, 40, 50, 60 and 75%, the short-term load-forecasting accuracy obviously decreases. In this research study, the short-term power-forecasting method is proposed in order to offset the power output of distributed PV power systems, and transfer to the original load without the effects of distributed PV power systems or treated as the input information for updated load-forecasting model, this study proposes the short-term power-forecasting algorithm based on the BP neutral network for the typical distributed PV power system through training and validation of the proposed model. The power-fore-casting accuracy is analysed and veri ﬁ ed for a typical distributed PV power system, and the proposed method is effective in improving the accuracy for load forecasting.


Introduction
The distributed photovoltaic (PV) power systems are generally connected with distribution power systems under the 10 kV voltage level in China. They are different from large-scale PV power plants, with regard to the capacity of the distributed PV power systems and local load demands. The power output of distributed PV power systems could be consumed by the local load or feedback distributed power systems. Consequently, the common algorithm of load forecasting is not suitable or less accurate, due to the uncorrelated effect factors that lead into the actual load variable considered the distributed PV power system as reversed load [1,2].
The short-term power-forecasting method is proposed in order to offset the power output of distributed PV power systems, and provide the actual load without the effects of distributed PV power systems. Various studies have been conducted on the shortterm distributed PV power system power forecasting; currently, the short-term power-forecasting method for distributed PV power system are categorised into two types. One type is the powerforecasting method based on solar radiation [3][4][5], this type is found in the detailed meteorological data, the forecasting process is complex for forecasting accuracy results. The other types directly forecast the day ahead power of a typical solar site, as omits the mathematic model building process.
Prior to the 1990s, the solar power-forecasting methods focused on the time-series method [6]. The rapid development of artificial intelligence technologies led to an advanced method which is used for PV power forecasting, such as support vector machine (SVM) and sparse Bayesian regression (SBR) [7][8][9][10]; however, the artificial intelligence technologies also have some drawbacks, and by using the SVM method it is hard to choose the core functions and parameters, while the SBR method is easy to use for part extreme values.
To analyse the effects of the distributed PV power system on common short-term load forecasting, 80 days of the actual load data of one 220 kV bus bar in the region distribution power system are collected. In addition, the relative power output and meteorological data of one typical distributed PV power system are analysed. Based on the proposed power penetration index and power and load data reconfiguration method, this paper simulated the accuracy variable trends of common short-term load-forecasting algorithm under the condition of the distributed PV power system power penetration as 0, 5, 10, 15, 20, 24.46, 30, 35, 40, 50, 60 and 75%. On the basis of the above research, it is apparent that the accuracy root mean square (RMS) of short-term load forecasting is gradually decreased. The short-term power-forecasting method is proposed in order to offset the power output of the distributed PV power systems, and transfer to the original load without the effects of distributed PV power systems or treated as the input information for updated load-forecasting model, this paper proposes the short-term power-forecasting algorithm based on the BP neutral network (NN) for the typical distributed PV power system, with the NWP models (especially ECMWF and Hebei Province Meteorological Bureau), based on the hidden layer transfer function as log-sigmoid function with ten layers. The power-forecasting accuracy is analysed and verified for typical distributed PV power system, and the proposed method is effective in improving the accuracy for load forecasting.

Analysis for PV power effects load forecasting
This paper collates power-generated data for a 40 MW distributed PV power system located in a city in the Hebei Province in People's Republic of China, that directly is connected to a 35 kV distribution network by transformers. Also, the 220 kV bus bar load data are collected from the substation where the distributed PV power system is connected to, the data are recorded as 5 min interval from 11 December 2016 to 28 February 2017 that lasts for 80 days. Fig. 1 is the original PV power data and 220 kV bus bar load data. In addition, the original load data, which is the sum of the PV power data and the actual load data, is also plotted.
The typical load data from 28 January to 31 January is plotted in Fig. 2, it is obvious that the actual load amplitude is changed due to the distributed PV power system generation, and due to this, the power output of the PV power system is lower when compared to the actual load, and the PV power generation is positive relative to the load, as a result, the variation of load data remains unchanged. On the side of the signal process, the load data signals are changed by amplitude, phase and attenuation, but not by the frequency.

Method for power and load data reconfiguration
To analyse the distributed PV power system effects short-term loadforecasting accuracy, this paper proposed the power penetration index that defines the penetration of the PV power for the distribution network as Based on the power penetration index, the actual power generated data and the load data obtained for variation in power penetration after analysis can be reconfigured. A detailed reconfiguration method for power and load data is shown from (2) to (8) where P lt is the 5 min load data and P l is the set for load data where P pvt is the 5 min power data and P pv is the set for power data where P elt is the 5 min original load data and P el the set for original load data where M pv ′ is the set for defined distributed PV power system penetration in distribution network where P ′ pv is the reconfiguration set for 15 min interval PV power as defined by the distributed PV power system penetration where P ′ l is the reconfiguration set for 15 min interval load data as defined by the distributed PV power system penetration.

Analysis of the effects of PV power
First, the load-forecasting model for 220 kV bus bar is required, the current used load-forecasting model is the NN model, based on  As shown in Fig. 3, the input layer transfers the information to the hidden layer, and the hidden layer transfers to the output, the input node is a no-calculation process and only takes values. Every node can have several inputs, but the output is unique and transfers to many nodes, as a result, forming the multi-layered front feed of the NN model. Especially, the node numbers, layers and cell numbers for the hidden layers are relative to the training results. The literature of NN used for load forecasting is common; this paper will not be dealt here.
Through the above load-forecasting model, and the reconfiguration distributed PV power system data and the load data, defined the penetration as 0, 5, 10, 15, 20, 24.46, 30, 35, 40, 50, 60, 70 and 75%, respectively, take the 50 days data for training the loadforecasting model and forecasts 30 January to 28 February; the short-term load-forecasting results are shown in Fig. 4.
From Fig. 4, it can be seen that mostly days, the higher distributed PV power penetration relative to the load-forecasting accuracy is worse. Especially, when the load-forecasting RMS is above 90%, which means that the PV power output has obvious effects on the normal load forecasting and higher penetration, the current loadforecasting results will be worse.
When the penetration is 0%, the current load-forecasting accuracy is 94.53%, when the penetration increases to 24.46%, which is current distributed PV power penetration, the load forecasting is 93.19%, which is 1.34% lower. Generally, the average of load-forecasting RMS is worse relative to the penetration of the distributed power especially when the penetration becomes higher (Fig. 5).

Short-term distributed PV power-forecasting model
This paper deals with the total solar radiation, diffuse radiation, ambient temperature and power data of the distributed PV power system from 26 January 2016 to 9 August 2016 for training and forecasting, the data interval is 5 min.

BP neutral network
As shown in Fig. 3, BP NN is a unique direction spread multi-layers forward network, the input for the hidden layer and the output is the weighting value for the output of the previous layer nodes, the excitation level depends on the excitation functions (Fig. 6).
Prepare the NN model with input X n and output Y n from any samples. The input of the number k node is given as where O j is the previous layer output and W jk the weighting value for the two layers. Also, the previous layers output and the excitation functions are as (10) and (11), respectively where u j is the bias and u 0 is used to adjust the shape of e function. The node numbers set affects the forecasting results, and the numbers are not able to express with an ideal equation. The node numbers are relative with the nodes of the input and output layers and accuracy is required. On the one hand, the node numbers are less, that will affect the obtained information for the network, and on the other, the training time will be longer and the error will  not be small. Therefore, the node numbers of hidden layers are important. This study ensures that the node numbers of the hidden layers (12) is 10 where n 1 is the node number, n the node number of the input layer, m the node number of the output layer and a the constant between 1 and 10.

Forecasting model training and validation
The excitation function of the hidden layer of BP NN is the logsigmoid function. Based on the data from the distributed PV power system, training the short-term power-forecasting model and the training results are shown in Figs. 7 and 8. Figs. 7 and 8 show that the excitation function and the training results fit well with the actual power output of the PV power system, the error rate of the training results is mostly around the region of ±3. Fig. 9 shows that R for the degree of linearity is above 0.98.
Based on the trained forecasting model, three types of input layer information are validation, as shown in Fig. 10, and by comparing the forecasting results mean squared error (MSE), the BP NN input layer information is total solar radiation, diffuse radiation, and ambient temperature.

Evaluation of power-forecasting results
On the basis of the built short-term distributed PV power system power-forecasting BP NN model, this paper developed the shortterm power forecasting from 19 March 2017 to 28 March 2017, while the period includes sunny and cloudy days, and the NWP meteorological data are obtained from ECMWF and the Hebei Province Meteorological Bureau for power forecasting. Fig. 11a shows the sunny day power-forecasting results; it can be found that the power-forecasting results curve is mostly the same with the actual power curve. Figs. 11b and c show cloudy days, relative to the sunny days, the cloudy days power-forecasting curve is bias from the actual power curve, the main reasons include the cloudy days training samples are less than the sunny days, and this is due to the variation in the position and thickeness of the clouds, which become hard to evaluate, the NWP solar radiation date is not accurate enough.
The typical day-ahead power forecasting and actual power output results are shown in Fig. 11, and the whole power-forecasting   Table 1.
Mostly, forecasting results accuracy is above 94%. However, due to the cloudy day effects, the power-forecasting results show the forecasting error as <91%. Generally, the built short-term distributed PV power-forecasting model accuracy meets the grid code requirement where the power forecast accuracy is 85%. With the increasing of samples and a detailed meteorological style, the power-forecasting accuracy will be better. In addition, with the development of the distributed PV power system, short-term powerforecasting technology, the forecast PV power can be used as the input data for the short-term bus bar load forecasting. Therefore, the actual load can be transferred to the original load, as a result, it will be useful for increasing the short-term load-forecasting accuracy especially for the high penetration distributed PV power connected distribution network.

Conclusion
On the basis of the proposed power penetration index and power and load data reconfiguration method, under the conditions of the distributed PV power system with different penetrations, the shortterm load-forecasting accuracy by the common load-forecasting model obviously decreases.
In this research study, the short-term power-forecasting method is proposed in order to offset the power output of the distributed PV power systems, and transfer the actual load demand to the original load demand without the effect of distributed PV power systems. This paper proposes the short-term power-forecasting algorithm based on the BP NN, and the proposed model is trained; in addition, the power-forecasting accuracy is analysed and verified for a typical distributed PV power system. The results indicate that the proposed method is effective in improving the accuracy for load forecasting.