The impact decades-long dependence on hydropower in El Niño impact-prone Zambia is having on carbon emissions through backup diesel generation

Emissions associated with hydropower are often forgotten. Lifecycle assessments of greenhouse gas emissions emanating from hydropower must count embedded carbon, emissions from reservoir lakes and the loss of carbon sinks, as well as backup diesel generation emissions when dependence on hydropower fails to deliver energy. Using Zambia as a case study, we estimate using a bottom-up approach that the emissions associated with backup diesel generation from Zambia’s power utility ZESCO and three largest sectors of consumers were up to 27 000 tonnes of CO2 in the worst months of drought in 2019. This is significantly higher than what a previous top-down approach would have estimated. We worked out ZESCO’s diesel generation attributable to drought using trend analysis. We worked out the mining sector’s emissions using copper production data, on-grid electricity consumption and calculated electricity intensity to infer off-grid electricity consumption in years of drought. From our household survey we learned average duration of generator use, average capacities of generators and acquired household income and generator use data which we ran in a Tobit regression. These together with labour force survey data helped us infer the level of diesel generation by households of different income brackets. For manufacturing firms we surveyed 123 firms. We collected rich diesel generation use data covering years of drought, input this into an OLS regression to identify predictors of diesel generation use (installed capacity of generator in kVA, in litres and whether generation was in a drought year) which we then used to extrapolate implied diesel generation for the firms for which we had less rich data. As global average temperatures and the frequency of El Niño droughts rise in hydropower dependent countries which account for a fifth of the world’s population, backup generation emissions have implications for the formulation of low carbon energy policy.


Introduction
Looking at proposed hydropower dams and the associated distribution infrastructure required in Chilean Patagonia, Mar (2009) forecast that the carbon impact of construction equipment and machinery, transport of labour, material embedded energy, and land-use change would be 48x the impact of natural gas plants that would deliver an equivalent amount of energy. Reservoirs themselves contribute to 1.5% of global anthropogenic emissions from carbon dioxide, methane, and nitrous oxide reported by the UN Intergovernmental Panel on Climate Change (Deemer et al 2016). On top of that, hydropower dependency can also lead to the release of greenhouse gas emissions in another way: when rainfall is low, hydropower ceases to be the dependable source of energy that the grid operator has used it for, and endconsumers (and even the power utility itself) resort to generating their own energy using diesel generators.
Farquharson et al pioneered thinking about the implications of electricity outages in sub-Saharan Africa for CO 2 , SO x and NO x emissions from diesel generators (Farquharson et al 2018). However, their top-down approach as well as their outdated data have underestimated the significance of emissions from backup diesel generation. This study aims to highlight the significance of backup diesel generation emissions in a hydropower dependent context using recent data and a bottom-up approach with Zambia as a case study.

Background
World Bank data (2020a) show that there were at least 58 countries as of 2015 that were more than 25% reliant on hydropower for domestic electricity production. 3 Among these countries, 33-whose combined population accounts for almost a fifth of the world's population-have been affected by droughts attributable to the El Niño effect. El Niño events have been forecast to increase in frequency with global warming (Wang et al 2017). This will necessitate the governments and power utilities of those countries, largely located in southern Africa, southeast Asia, Australasia, and northern South America (FAO 2015, Vidal 2016, Hao et al 2018, to think about how to make their energy generation infrastructure more climate resilient. It will also necessitate thinking on how to avoid contributing to the global warming that is making their hydropower infrastructure increasingly redundant. We show that unreliable hydropower infrastructure paradoxically results in increased purchases and use of high-greenhouse gas emitting diesel generators.
We took Zambia as our case study. As of 2015, Zambia was one of at least 12 countries, of which 11 were not high-income countries, that were more than 80% reliant on hydropower for their electricity production (Zambia was at that time 97% dependent), while at least 35 countries, of which 26 were not highincome, were more than 50% reliant on hydropower (World Bank 2020a, 2020b.
For decades, using hydropower as a dispatchable source of energy was not the reason why Zambia's power utility ZESCO sometimes failed to deliver energy reliably, since energy supply with properly maintained infrastructure would have far outstripped demand. But then: 1. Demand caught up with supply. By 2015 electricity demand was 1959 MW, compared with installed capacity of 2411 MW (email sent by an energy officer at the Ministry of Energy on 30 April 2020); average capacity only needed to fall below 81% for Zambia to experience power outages.
2. Energy supply dipped. Low levels of rainfall (World Bank 2018b) resulted in lower reservoir levels in Zambia's hydropower dams (interviews with ZESCO Ltd staff, 10 November, 2017 and 6 June, 2018). The result was load shedding lasted a minimum of 8 h a day for the majority of ZESCO's customers (Energy Regulation Board of Zambia 2016, p 1). Less energy was consumed in Zambia in 2015 than was consumed in 2014 and less energy was consumed in Zambia in 2016 than was consumed in 2015 (Energy Regulation Board of Zambia 2016, Zambia 2017 Zambia 2017, p 15). The effect was that electricity-intensive manufacturing, a traditional catalyst for industrial-led economic growth (Szirmai andVerspagen 2015, Cantore et al 2017), lost not only production potential in 2015 and 2016-when its absolute electricity consumption fell below its 2014 level-but that by 2017 it had also lost relative electricity consumption share to finance and property (see table 1).
Had ZESCO not mitigated its generation shortfall with additional diesel generation and more importantly the import from the Southern African Power Pool (SAPP) of 785.2 GWh in 2015 (up from just 12.8 GWh in 2014) and 2184.9 GWh in 2016 (ibid, p9), the deficits would have been even greater. Grid electricity consumers also turned to alternative generation strategies. Figure 1 below illustrates Zambia's electricity and electricity substitute systems boundary. It builds upon Bayliss and Pollen's system of provision (Bayliss and Pollen 2019, figure 1) which was limited to grid electricity consumption. This system includes the 77% of the population that live without grid connections (ZICTA and CSO 2018) and who rely on off-grid solutions including solar products and biomass (Tembo 2018) 50 years after the World Bank's substantial energy investments into Zambia (World Bank 2018a) to support mining IDA 1970, 1973) in the vain hope that the benefits of hydropower would trickle down (Park 1968). Personal interviews with Africa GreenCo (4 June 2018, 23 April 2019, 3 September 2019) revealed that the company intends to facilitate investment into renewable Independent Power Projects by becoming a private credit-worthy off-taker that on-sells to ZESCO, to clients on the SAPP and to bilateral clients. Agricultural firms can meet their electricity needs using their biomass (Sugar 2017, p. 15). Manufacturing firms use backup diesel generation to make up for lost ZESCO power (Ahmed et al 2019), as do mining companies (Mfula 2010a, 2010b, News 2011, Syndigate 2011. Grid-connected households' favourite mitigation strategy to power outages is rechargeable lights and solar products; 10th and 11th favoured strategies are UPSs and diesel generators (Ahmed 2020). Our earlier findings that manufacturing firms and households self-generate energy using diesel to the extent that they do (Ahmed et al 2019, Ahmed 2020) motivated this study whose aim is to quantify and aggregate the CO 2 emissions from diesel generation by ZESCO as well as Zambia's three largest types of electricity customer at the advent of its first set of power outages in 2015: mining companies, households and manufacturing firms.

Methods
We adapted our approach according to the homogeneity of a sector and, for the sectors where we collected data from respondents, the granularity and verifiability of data available. The Energy Regulation Board of Zambia provided diesel generator use by the power utility ZESCO. Using trend analysis we worked out the difference of ZESCO's diesel generator use between years when there was no load shedding and years in which there was load shedding due to low hydropower reservoir levels. Because mining is a fairly homogenous sector (explained further below), we relied on existing secondary data from the Energy Regulation Board for the sector's on-grid electricity consumption and the US Geological Survey for production and broad assumptions about production (see table 2 above). Broad assumptions would have been unsafe for households and the manufacturing sectors which are a lot more diverse. For this reason, we collected primary data for these sectors. Nine manufacturing firms recorded the litres of diesel fuel that they consumed, installed capacity of their generators and the hours that they ran their generators for, while 84 of the 91 firms that reported using self-generation that we interviewed provided some of this data. We were thus able to extrapolate fuel use by the sector using OLS regression analysis with a strong degree of confidence. For households we found income to be a statistically significant predictor of generator use and ownership, and we therefore extrapolated fuel use using this relationship.

ZESCO's diesel generation
To make up for the energy shortfall during the power outages of 2015 and 2016, ZESCO resorted to generating energy from its diesel power plants. We used Energy Regulation Board reports of ZESCO backup diesel generation use from 2008 to the most recently available year to work out changes in diesel generation, which in turn helped us to work out the extent of diesel generation in 2015 and 2016 that was attributable to the hydropower dams not being able to generate sufficient energy in those years.
We calculated diesel generation attributable to drought over a year by ZESCO (ZESCO DGD) as where DGYD is diesel power generation in year of drought, GWh.  And DGPD is diesel power generation in last year prior to drought, taking into account growth of use of diesel generation, GWh.
From diesel generated energy attributable to the drought, we worked out the implied CO 2 emissions using publicly available conversion rates.

The mining sector
Although mining is by far the largest electricity consumer in Zambia, we anticipated that it would also be the least diverse in terms of its energy use patterns among ZESCO's clients. Most of Zambia's mining is for copper 4 , and 80% of Zambia's copper comes from just four mines (Zambia Chamber of Mines 2016).  Both open pit and underground mines use diesel generation to prevent ZESCO power outages from resulting in loss of production (Mfula 2010a, 2010b, News 2011, Syndigate 2011, although underground mines are more electricity intense (Tembo 2018, p 41). Use of generators by the mining sector preceded the El Niño droughts of 2015 and 2016; mines experienced power outages for various reasons, ranging from apparent vandalism to fires damaging ZESCO equipment (Jourdan 1990, Mfula 2010a (also email dated 17 February 2019 from Michael Mainelli, assigned to advise on the privatisation of Zambia Consolidated Copper Mines in 1992).
To calculate the emissions emanating from the sector, we looked to the average historic energy intensity associated with mining production in years prior to 2015 to impute what level of electricity production (both on and off-grid) would have been required to extract that which was mined in 2015 and 2016. The difference between this value and the ongrid energy consumption of the sector would tell us what the sector likely consumed from off-grid backup diesel generation.
We calculated diesel generation attributable to drought over a year by the mining sector (Mining DGD) as where Mining EI is average electricity intensity of mined copper, GWh/'000 tonnes.
Cu is copper mined over the year, '000 tones.
Mining GE is grid electricity consumed by the mining sector over the year, GWh.
From diesel generated energy attributable to the drought, we worked out the implied CO 2 emissions using publicly available conversion rates.

Households
We now turn to the second largest consuming sector of electricity in Zambia: households. The Zambia Information and Communications Technology Authority and Zambia's Central Statistical Office estimated that in 2018, Zambia had a population of 16.9 million people across 3.5 million households, grid energy access of 33%, and internet access of 18% (ZICTA and CSO 2018). This implies that 630 000 households had internet access.
After obtaining ethical approval from University College London, we collected surveys from 54 individuals online in December 2019. 6 We shared the survey with our personal contacts by email, What-sApp, LinkedIn and Twitter, and requested these contacts to share the survey with their contacts. Although 128 surveys were initiated, the majority did not sign the ethical permissions required to proceed with answering the survey questions. We stopped the survey within a month because of a regulatory 6 The data that support the findings of this study are available at https://imadahmed.com/2020/07/26/erl-diesel-emissions-data/. Full survey data and questionnaires will be made available upon the publication of another journal article and Imad Ahmed's PhD. change in tariffs which we believed would contaminate the results of the rest of our survey questions. 83% of respondents answered from Lusaka Province and 9% from Copperbelt Province. There was one respondent (1.8% of respondents) for each of four other provinces: Muchinga Province, Easter Province, Southern Province and Luapula Province. The sample collected therefore overrepresented Lusaka by 2.2x in terms of active internet users, underrepresented the Copperbelt by 1.7x, and underrepresented Southern Province by over 30 times. Together, these three provinces accounted for 71% of Zambia's active internet users (ZICTA and CSO 2018, figure 79). Had data existed on the total residential-use data purchased disaggregated by province, the sampling achieved may not have been so over-weighted for Lusaka.
Each respondent used a distinct internet protocol address. Between this and the fact that no two set of responses was alike, it is likely that each response represented a distinct household. We asked respondents whether they owned generators, to what extent they used generators, their income brackets, and their ages. Where we found respondents to be aged 25 or less and in the lowest income brackets, we discarded their responses as those who were 25 or younger and in the lowest income bracket and were using generators may have been doing so with support from their parents, and so should have been counted in their parents' households, or were getting the benefit of diesel generators thanks to their university accommodation-the survey had reached University of Zambia students. These under 25 year olds would not have been representative of the Zambian population. We cleaned the data for these 25-year olds, who accounted for 7 of the 54 responses.
We calculated diesel generation over a month by households following these steps: 1. We asked households which monthly income bracket they fell into and how much they used their generators on a scale of 0 to 4. 7 2. We ran a Tobit regression with generator use as the dependent variable and household income as the independent variable. 3. To get relative generator use intensity for each household income bracket (RI), we a. multiplied the statistically significant coefficient by each household income bracket b. divided each intensity score by the highest intensity score (scored by the highest income bracket since the coefficient was positive). 4. We then calculated diesel generation by households over a month (HH DG) as Where IB is income bracket HIB is highest income bracket LIB is lowest income bracket for which data was collected %UG IB is use of generators per household income group #HH IB is number of households in income bracket RI IB is relative generator use intensity for each household income bracket DU is average daily use, hours C is average generator capacity, translated into kW 30 is average days/month From diesel generated energy, we worked out the implied CO 2 emissions using publicly available conversion rates.

Manufacturing firms
Unlike copper mining which is dominated by four mining companies and mining pits in Zambia, the manufacturing sector is diverse. In the capital Lusaka, the Ministry of Trade and Commerce noted over 80 types of manufacturers (Ministry of Commerce Trade and Industry 2014, pp. 133-139). Load shedding in Zambia seems to have different impacts on different manufacturing subsectors (Sichone et al 2016); diesel generator use also seems to vary across Africa by subsector (Steinbuks and Foster 2010).
To collect firm-level data on generator usage, we hired a team of enumerators through funding made available by the International Growth Centre and managed to survey 123 firms between April-August 2018 in Zambia's largest manufacturing hubs: Lusaka, and Ndola and Kitwe in the Copperbelt Region. We wanted to collect data on capacity of generators used in terms of kVA and litres, usage data varying between drought and non-drought years and varying with manufacturing subsector, year of purchase of first generator, losses experienced as a result of power outages and the extent to which firms successfully mitigated these (Ahmed et al 2019). We aimed to survey a representative sample of Zambia's large manufacturing firms by subsector and by geographic industrial hub. To do this, we studied the national breakdown of manufacturing subsectors according to the latest census data we were able to find (Ministry of Commerce Trade and Industry 2014). In this regard, we succeeded. In our sample, no subsector was over or underrepresented by more than 9% (see table 4). With regards to sampling by representative geographic hubs, we under surveyed in Kitwe and Ndola in the Copperbelt Province. We aimed to have 59% of the sample come from Lusaka and the remaining from the Copperbelt (since 50% of large manufacturing firms are in Lusaka and 84% of large firms are in both Lusaka and the Copperbelt; 50/84.1 = 59%). In fact, 71% of the sample came from Lusaka.
We calculated CO 2 emissions emanating from diesel generation over months in drought years and months in non-drought years by manufacturing firms following these steps: 1. Nine firms either provided information of MJ of energy produced by diesel generation/month or litres of diesel consumed in power generation over the years 2015-2018. For the firms that provided information on litres of energy combusted, we converted these into energy values using publicly available conversion factors for litres of diesel combusted into energy. 8 We then converted energy yielded from diesel generation into CO 2 emissions using publicly available conversion factors for energy generated from diesel generation. 9 2. These nine firms also provided their subsector, kVA of installed diesel generator capacity and 6 provided their installed diesel generator capacity in litres. We therefore ran ordinary least squares regression with CO 2 emissions as the dependent variable and the following variables as independent variables: CO 2 emissions = a * year of drought 10 + b * installed capacity expressed in kVA + c * installed capacity expressed in litres + d * whether the firm is in the food subsector + constant.
3. We then multiplied the regression coefficients by the characteristics of the remaining firms, 91 of which had generators, of which 84 provided kVA installed diesel generator capacity values and 41 provided litres of installed capacity. All firms provided information on their subsector. Table 5 row 1 shows the total energy generation by ZESCO using diesel. Energy generation using diesel increased by a compounded annual growth rate of 10% (see row 3) from 2010 to 2014, although growth from 2013 to 2014 of 3% (see row 2) suggests that it was plateauing. If we ignored this plateauing and used the CAGR for 2010 to 2014, row 5 shows the implied excess diesel generation attributable to low hydropower reservoir levels in 2015 and 2016; row 6 shows the same but in TJ and row 7 converts this into carbon dioxide emissions.

ZESCO's energy generation using diesel
Taking into account the plateauing, if we assume ZESCO's diesel generation would not have increased past its 2014 consumption level but for the drought that effected power outages in 2015 and 2016, row 8 shows what the energy produced by diesel generation would have been, row 9 calculates the implied excess attributable to low reservoir levels, row 10 shows the same in TJ and row 11 converts this into carbon dioxide emissions.
Due to low hydropower dam reservoirs in 2015 and 2016, the emissions emanating from increased used of diesel generators by ZESCO could have been as high as 1230 tonnes of CO 2 (100 tonnes/month).

The mining sector
Row 3 of table 6 bears out our prediction that Zambia's mining energy use patterns are predictable. For each of 2013, 2016, 2017 and 2018, mining's gridelectricity intensity was 7.8 GWh/1000 tonnes of copper.
Using this average grid electricity intensity with the copper produced for each of the years, we computed the predicted grid electricity consumption for each year (row 5). Where the predicted grid electricity consumption was greater than the actual grid electricity consumed, the variance was noted in row 6. We see a great deal of variance in 2016-the second year of power outages in Zambia.
We can plausibly infer that Zambia's mining sector self-generated 47 GWh of energy in 2016, and that up to 100% was generated from diesel generators. Zambia's mining sector would in an average month in 2016 have contributed to 1000 tonnes of carbon dioxide. 11

Households
Two-thirds of households in the highest income bracket reported using diesel generators when grid power went out, 30% of households reported using diesel generators in the next highest income bracket (see figure 2 below).
Five households that used generators provided information on their installed capacity in kVA, the median of which was 5kVA. At a power factor of 0.8, this is the equivalent of 4 kW. Owners generally used their generators after sunset (6pm) for an average of 5 h. 12 This implies that the respondents used their generators until they slept at approximately 11pm. Our sample of installed kVA and hours of generation use was not large enough to regress against household income given the limited richness in data. However, our sample of 47 households that reported both their 11 1 GWh of energy is equivalent to 3.6 TJ of energy (IEA 2019) and there are circa 74 100 kg of carbon dioxide for every TJ of energy produced from diesel (Gómez et al 2006). 12 Treating the 20 h response as an outlier, the average of five responses was 4 h of a power outage before respondents switched on their generators. The average duration of five responses was 5.2 h. Table 6. Imputing diesel-generated electricity by the mining sector.  , p 8, 2016, p 0, 2017, p 9, 2018, p 36, 2019, p 39) and U.S. Geological Survey (2015, 2020.   7). Because our dependent variable was ordinal, we used Tobit regression analysis on income as the independent variable, which was statistically significant and had a positive effect on generator use at the 5% level. The Chi-square of the model was also statistically significant at the 5% level as shown in table 8. The interpretation of the coefficient was that every additional kwacha of income earned per household per month resulted in extra use of the diesel generator by 0.0000254 units on the scale of 0 to 4. This means that the Tobit regression model predicts that the highest income bracket used generators on average 1.2 ('use a bit') and the second highest income households were using them on average 0.7 (less than 'a bit') contrasting with the average usage of 1.7 for the highest income bracket and 0.8 for the second highest income bracket (see figure Coefficients first, standard errors in parentheses. * * * p < 0.01, * * p < 0.05, * p < 0.1.

Row
2 above). What this effectively means is that households in the highest income bracket were using generators 160% more than households in the second highest income bracket, in which we can expect to find the median installed capacity of power generation and median duration of hours of generator use (see the third column of table 9). 13 Households in the next income bracket use generators diesel generation 35% of the level that households in the second highest income bracket do. These results are shown in columns 5 and 6 of table 9 below. Household carbon dioxide emissions from diesel generation were approximately 2900 tonnes in the worst months of 2019.

Manufacturing firms
We found that more than 73% of firms had acquired use of a backup generator, and that 52% of these had bought their oldest generators in the years of Zambia's worst power outages in 2015 and 2016 (see figure 3).
If the age of the oldest generators on firms' premises can be taken as a proxy for the first time that firms bought their generators, then the data on which Farquharson et al (2018) based their topdown analysis of generator emissions for Zambia, at least, was out of date because it was collected by the World Bank from December 2012-February 2014 (World Bank and International Finance Corporation 2014). The purchase year of firms' oldest generators was directly observable by enumerators and did not rely on institutional memory which is vulnerable to change in personnel and wrongly remembered events.
For the years 2014 to 2018, we asked respondents to rank them on a scale of 0 to 4 for losses. 0 represented no losses in production, 1 represented 1%-15% losses of targeted production, 2 represented 16-30%, 3 represented 31%-50% losses and 4 represented more than 50% of losses. To account for the greater number of responses for the most recent years, the aggregate score for a given year was divided by the number of responses for that year. 2016 and 2015 ranked as the worst years (see table 10). Surviving firms reported losses towards the higher spectrum of 16%-30% in 2015 and 2016.
Supporting our finding that more than 80% of firms that saw reduced revenues as a result of power outages is the 97% correlation between the average scores for losses for 2014 to 2018 with the number of oldest generators firms had that were purchased in the years 2014 to 2018 (see figure 4). 93% of firms that had had access to a generator and whose generator did not break down continued to use their generator.
From nine firms that provided detailed accounts of the extent of energy then had generated from diesel generators or litres of diesel they had consumed in generating energy on a monthly basis over 2015-2018, we calculated the emissions associated with  their diesel generation and ran these as dependent values in an OLS regression against four independent variables: installed generator capacities in kVA and litres, the years in which they ran their generators and their subsectors. The first three were statistically significant predictors. If the diesel generator was used in a drought year, CO 2 emissions for the month would be 45 000 kg greater. Each additional kVA of installed capacity translated into 18 kg of additional CO 2 emitted and each additional litre of installed capacity translated into an additional 17 kg of CO 2 . We did not find statistical significance for various subsectors in our sample size which limited the number of subsectors that were observed (most belonged to the food and beverages subsector). Nevertheless, the Rsquared of the model showed that the OLS regression model explained 99.9% variance in the dependent variable and because the p-value of the associated Fvalue of the model is less than 5%, we can say that the independent variables reliably predict the dependent variable.
Using the coefficients in table 12 above, we predicted emissions for Zambian manufacturing firms   for which we collected data on the level of installed capacity of backup generation in terms of kVA and litres, as explained in the methodology above. Extrapolating the results for 123 such firms, the sum of CO 2 emissions from backup generation came out to   Zambia ( , 2016Zambia ( , 2017Zambia ( , 2018 representative of the population (which table 4  As noted earlier, the number of daily hours of outages apparently doubled for load shedding from 8 in 2015 to 15 at the end of 2019 (Energy Regulation Board of Zambia 2016, Zesco 2019. It is therefore probable that mining firms and manufacturing firms would have been even worse hit with power outages in 2019 during their hours of operation. Depending on their ability to finance extra diesel generation and how squarely the outages fell during their hours of operation, emissions could have increased by up to 100%. The range for CO 2 emissions from diesel generators from these three sectors at the end of 2019 from Zambia's power outages could therefore have been as high as 27 000 tonnes of CO 2 /month. We will be able to calculate the emissions from ZESCO's diesel power plants in 2019 when the Energy Regulation Board publishes its report for 2019.

Uncertainty analysis
Finance and property overtook the manufacturing subsector as the third largest consumer sector of ZESCO electricity in 2016. We were unable to estimate backup diesel generation emissions for this sector. We do not expect it to have contributed to emissions to the same extent as energy-intensive manufacturing firms would have, but would have expected this sector to have been more energyintense than ZESCO's high-income households since client-focused professional service sector offices cannot afford to not have power outages during office hours.
Given that new power plants were also coming online for ZESCO and the fluctuating capacity utilisation rates for its existing hydropower assets, we also did not feel confident in predicting how much energy shortfall ZESCO would have tried to make up using diesel generation in 2019.
There was significant sensitivity (resulting in 26 000 tonnes/month for the aggregate result versus 27 000 tonnes/month) in the currency conversions and salary escalations we assumed from 2012 Zambian kwacha to December 2019 United States dollars for the purpose of figuring out how many households existed in each of the higher middle class income brackets according to Central Statistical Office Labour Force Survey data (Zambia Institute for Policy Analysis & Research 2013).

Conclusion
As noted earlier, Farquharson et al (2018) used out-of-date data for their model. For Zambia, they estimated that a typical month saw 5.2 outages for an average of 2.8 months; that the installed grid capacity was 2.3 GW and that generator availability was 3% of the grid. This would imply that they estimated that generators on average produced 1 GWh of energy in a month, the equivalent of 3.6TJ, resulting in just 74 tonnes of CO 2 . This is not close to the same level of magnitude of emissions that we estimated for just one sector of the economy.
To triangulate our results, therefore, we need to perform our own estimates using more realistic highlevel data for a back of the envelope ratification that Farquharson et al's study does not provide. 2018 saw 13 000 GWh of grid electricity consumed in Zambia (see table 1). A twelfth of that is 1100 GWh in a month. 15 h a day of outages would imply that most of a work day or an evening at home postsunset would require a generator for the 33% of Zambia's population that enjoy on-grid connections. If we assumed that generators covered 15% of grid capacity for a country with an established mining industry that consumes more than half of grid-electricity, and that generators were used for half the time that there were power outages, CO 2 emissions would be 22 000 tonnes. By this calculation, our bottom-up approach estimate seems reasonable.
Our bottom-up analysis suggests that the manufacturing sector uses diesel generators more intensively than the other sectors that consume more energy than it does. From ZESCO's load shedding schedule of October 2019, we can see that ZESCO is attempting to minimise disruption to manufacturing firms during business hours (Zesco 2019). This is a good idea. ZESCO is also installing increased power generation capacity (Zesco Ltd 2017b). However it is doing this through hydropower which will continue to fail in times of drought and through emissions-intensive coal-fired power. We have a couple of alternative suggestions for ZESCO.
First, ZESCO should consider prioritising clients that use backup diesel generation the most for more reliable energy. To achieve this end, ZESCO should look into the feasibility of sourcing these clients' energy from the SAPP's nuclear thermal station.
Second, ZESCO should seek to dampen inefficient energy use. It could do this by increasing average tariffs across the board, which would also help it achieve system sustainability so that it can invest in increased installed power generation capacity or afford the cost-recovery tariffs charged by energy exporters on the SAPP. Beyond increasing tariffs across the board, ZESCO could also increase specific tariffs. It could increase the peak tariffs to decrease peak demand. Research by Ahmed et al (2019) showed that manufacturing firms use more peak-hour energy than predicted because peak energy was under-priced. Increasing the price of peak energy would reduce demand for peak energy, if the tariff is increased sufficiently so that demand elasticity becomes negative against price increases. Another mechanism for dampening inefficient energy use would be for ZESCO to charge a penalty for reactive power.
Third, ZESCO and the Energy Regulation Board should investigate ZESCO's elevated sources of distribution and transmission losses and work to address these before commissioning extra power supply. ZESCO's transmission and distribution losses were recorded as 10% (Energy Regulation Board of Zambia 2017, p 63) according to self-reported data by ZESCO. An Energy Regulation Board official noted that this was too good for the kind of data his agency was able to extract from what they were given by ZESCO. The implied losses of generated energy grew to almost a fifth of energy generated in 2016, calculated by subtracting from the output generated the net exports and domestic consumption and dividing that by the output generated (see table 14 below)-well above the 12% threshold below which ZESCO is required to maintain distribution losses. This could be the result of both transmission losses and non-technical losses, such as pilfering and unmetered use. ZESCO economists said that these calculations in line 6 of the below table were credible calculations (personal communication, 3 September, 2019, ZESCO Ltd, Lusaka). The transmission and distribution losses in 2015 and 2016 were greater than the difference in domestic energy consumption between 2014 and 2015 and 2014 and 2016 respectively.
In terms of reliable supply to meet Zambia's demonstrated demand for grid electricity, crosscontinental super-grids could be the answer. With the right governance structures in place to prevent holdups and renegotiations, they would allow the transfer of energy from jurisdictions with surplus lowgreenhouse gas energy generation infrastructure to jurisdictions with deficient energy generation infrastructure. Within the SAPP, the Koelberg nuclear power station in South Africa provides dependable baseload energy that is emissions-free during operation, but its capacity is insufficient to power all of southern Africa and reports of corruption behind the additional procurement of nuclear power (Cornish 2019) have not been confidence inspiring. Namibia has enormous solar potential, though without better battery storage, intermittency will remain an issue. Tanzania and Mozambique's natural gas may do almost the same harm as Zambia's coal (BEIS 2019, table 4) for climate change when well-to-tank methane leakages are taken into account (Marchese and Zimmerle 2018) but the immediate effects of their emissions on human health would be less deleterious. Beyond the SAPP and beyond Africa's intermittent or climate vulnerable renewable energy potential are France's nuclear power plants.
To the extent that Zambia's demonstrated demand for grid electricity is an under representation of demand by those without grid connections, and to the extent that energy cannot be so easily transferred across continents to prevent emissions from the construction of power plants and the destruction of carbon sinks, the next solution requires even greater political will. If populations privileged with power want to prevent a proliferation of fossil-fuel plants in hydropower dependent countries falling prey to anthropogenic climate change-induced drought, they should welcome the orderly immigration of people from those countries.
The case of reduced hydropower capacity in Zambia is unfortunately not a unique one. Up to a fifth of the global population live in nations heavily dependent on hydropower that is vulnerable to drought due to increases in average global temperatures. The above suggestions for how Zambia can overcome its energy challenges in a low greenhouse gas intensive manner could therefore have broader applicability.