Impact of Climate Change on Residential Electricity in Western Cities in China

Extreme weather induced by climate change has triggered large-scale power outages worldwide. More insight into the climate impact (especially precipitation) on residential electricity consumption (REC) is needed. With a particular focus on precipitation, our study aimed to quantify the climate impact on REC and project associated changes under the representative concentration pathways (RCPs) 2.6, 4.5, and 8.5 climate change scenarios in Lanzhou and Lhasa, two western cities in China. The climate impact on REC in both cities is driven by heating-related demand, with Lanzhou being more sensitive to climate than Lhasa during the warm season. There is a stronger precipitation impact during the cold season in both cities, since precipitation (especially snowfall) boosts electricity consumption during the cold season. As the two cities become warmer and wetter, precipitation will offset the impact of warming on REC, with Lanzhou being more affected. Based on Lanzhou projections, the offsetting effect will be more pronounced in the cold season across all scenarios, and will be particularly strong under RCP 2.6. For the remainder of the year, the effects of increased precipitation and warming will have competing importance under the RCP 4.5 scenario, whereas temperature effects will generally dominate the climate impact under the RCP 8.5 scenario. Our results provide new references for future climate–energy studies in western China and can help inform policy-making on regional climate adaptation. We also propose an algorithm model that is readily compatible with potential early-warning projects to protect against extreme weather-induced power outages.


Introduction
Extreme weather events induced by climate change threaten energy supplies worldwide (Mideksa & Kallbekken, 2010). Understanding how residential electricity consumption (REC) responds to different climatic conditions as well as socioeconomic conditions is key to mitigating these threats, as it will assist in evaluating the social cost of carbon, i.e., a key indicator for policies focused on curbing climate change (Li,  however, there are fewer data points for precipitation than for temperature for a given study period (Davis & Gertler, 2015;Deschênes & Greenstone, 2011). An alternative is to treat precipitation as a covariate or control variable (Auffhammer et al., 2017;Zhang et al., 2020), which effectively means paying less attention to precipitation than to temperature. However, precipitation doesaffect REC (Auffhammer & Aroonruengsawat, 2011; Deschênes & Greenstone, 2011), and precipitation changes were shown to increase the predictive uncertainty of REC in an eastern province in China .
Our study aimed to quantify the impact of climate on REC and project its future trends under climate change in two western cities in China, with a particular focus on the role of precipitation. We hypothesized that precipitation impacts heating-related REC in the cold season, likely in the form of snowfall. If our hypothesis is correct, climate impact on the REC in the cold season may not simply decline with warming; it may also increase if there is more precipitation in the cold season. To test this hypothesis, we built an algorithm model using seven machine learning algorithms and predicted future climate impact on REC under three climate futures projected with different greenhouse gas concentration pathways (RCP). They are RCP 2.6, i.e., whereby greenhouse gas emissions are expected to start to decrease by 2020 and reach zero by 2100; RCP 4.5, whereby the emission is assumed to peak around 2040 and then decline; and RCP 8.5, where the emissions continue to rise throughout the 21st century.
We selected two western cities in China,Lanzhou and Lhasa, as our research objects. The two cities were chosen to represent different climate pattern in western China. Lanzhou is located in the northwest and Lhasa is in the southwest. We also selected these two cities as they represent the different heating infrastructures in northern and southern China, as de ned by the Qinling-Huaihe Line. Lanzhou is equipped with a coal-fueled central heating system as in other northern cities, while Lhasa relies mostly on household electric heating as inthe rest of southern China.
By investing climate-REC relationships in these two cities and project the impact of climate change, our results provided new references for future climate-energy studies in western China and can help inform policy-making on regional climate adaptation. In the process, we also developed an algorithm model that is readily compatible with potential early-warning projects to protect against extreme weather-induced power outages. and urbanization. Since Lanzhou and Lhasa are both cities, we considered information on urbanization incorporated in the population data. Thus, we built our model using temperature, precipitation, income, and population as predictor variables, and REC as the response variable. Unlike previous studies using data modeling, we adopted an algorithm modeling approach. With respect to the predictive accuracy as well as interpretability, we compared the results from our model with those from the previously used parameter-based models to verify and demonstrate the strength of our new modeling approach. We then tapped into the capacity ofour model to holistically estimate climate impact by projecting the impact of climate change on REC under RCP2.6, RCP4.5, and RCP8.5 scenarios.

Data
We based our estimates on 2011-2019 Lanzhou and 2014-2019 Lhasa panel data. Each dataset included monthly REC, monthly mean temperature, monthly precipitation, annual salary, and annual population. The the study period was mainly de ned by the availability of monthly REC data for the two cities. We obtained Lanzhou REC data from the Gansu State Grid Company (GSGC) and Lhasa REC data from the Tibet State Grid Company (TSGC). Monthly mean temperature and monthly precipitation data were calculated using 2011-2019 Lanzhou and 2014-2019 Lhasa daily temperature and precipitation data from the China Meteorological Administration. Yearly salary and population data were provided by the National Bureau of Statistics of China.
For the prediction of future climate impact on REC in the two cities, we used future climate projection of China based on the regcm4.6 model under the RCP2.6, RCP4.5, and RCP8.5 scenarios (Lei & Xiaoduo, 2020). This regional climate model has yielded more accurate simulation results of present-day mean climatology over Northwest China than the Met O ce Hadley Centre Earth System (HadGEM2-ES) global climate model (Pan, Zhang, & Huang, 2020). We extract climate projections for the two cities from the nearest available points in the data; i.e., at the coordinates of 36°N,103.5° E for Lanzhou and 29.5°N,91° E for Lhasa. This climate projection data is more reliable for Lanzhou, as it is located in Northwest China where the regcm4.6 model performs well. Lhasa, on the other hand, is located not only in Southwest China, but also in the Tibetan Plateau. While the Tibetan Plateau is found to be getting warmer and wetter (Yao et al., 2019), climate projections for the region are greatly complicated by the uplift of the Tibetan Plateau and there is not reliable near-term climate change projection yet (Hu & Zhou, 2021). Since Lanzhou is also projected to get warmer and have increased precipitation (Lei & Xiaoduo, 2020), we were able to obtain a qualitative reference for Lhasa from results in Lanzhou.

Algorithm modeling vs. data modeling
We adopted algorithm modeling rather than data modeling as used in previous studies. Modeling can be considered as a process associating predictor variables with the response variable; the main difference between data modeling and algorithm modeling is the way in which the association is built (Leo Breiman, 2001). Data modeling assumes that real-world data conform to a stochastic regression function that incorporates predictor variables, random noise, and parameters, and attempts to estimate the response variable using functions that are then validated using goodness-of-t tests(Leo Breiman, 2001).
Algorithm modeling, on the other hand, assumes that the association is unknown and attempts to identify an algorithm that can best predict the response variables using predictor variables, and the results are then validated using predictive accuracy (Leo Breiman, 2001).
Data models attempt to t data gathered from real-world observations of various systems with unknown mechanisms into a parametric model selected by a statistician. This inevitably compromises accuracy as well as insight (Leo Breiman, 2001). Despite efforts to minimize parametric assumptions (Deschênes & Greenstone, 2011), the econometric models used in previous studies remain as parameter-based data models with similar limitations. Conversely, algorithm modeling is completely data-driven since it makes only one assumption, i.e., that the used data are obtained from an unknown multivariate distribution (Leo Breiman, 2001). The more accurate a model, the more reliable is the information we can derive regarding the underlying data mechanism. Since algorithmic models have generally been proven to make better predictions (Leo Breiman, 2001), we adopted algorithm modeling in our study.
Since algorithm modeling treats the association between predictor and response variables as unknown, there was no means of ensuring that the algorithm worked well with our data prior to testing. We derived a pool of seven commonly used candidate algorithms (Brownlee, 2016).
Three are basic algorithms: k-nearest neighbors (KNN), support vector regression (SVR), and classi cation and regression tree (CART). KNN was designed with the assumption that similar things are in close proximity to each other. It makes estimations by measuring the distances between an instance and all other instances in the dataset. It selects the speci ed number of instances (K) closest to the one concerned, then votes for the most frequent label (in the case of classi cation) or averages the labels (inthe case of regression) (Altman, 1992). SVR works by constructing a hyperplane or hyperplanes in a high-dimensional space for either classi cation or regression. The hyperplane with the longest distance from all the nearest training data points of any class (also known as the functional margin) is considered a good separation since the larger the margin, the lower the generalization error (Cortes & Vapnik, 1995).
CART is also known as a decision tree, since it is a tree-like algorithm that is built through repeatedly splitting into two child nodes. It splits on a single input variable that improves the Gini index, a performance measure that calculates the probability of a speci c input variable being classi ed incorrectly when selected randomly (L Breiman, Friedman, Olshen, & Stone, 2017). The structure of CART makes it (along with tree-based algorithms) innately immune to correlation among predictor variables (J. The other four are ensemble algorithms. Ensemble is a machine learning method that combines several base algorithms into one in order to decrease variance and bias or to improve predictions. We used Adaptive Boosting (AB), Random Forest (RF), Extra Trees (ET), and Gradient Boosting Machine (GBM), i.e., all CART ensembles with different ensemble methods . RF and ET train their individual trees independently and average predictions across trees, an ensemble method known as bagging. RF and ET are different in how their individual trees split as well as in how they sample data. RF splits where the performance measure is best, whereas ET splits randomly. RF subsamples the data sample with replacement, or bootstrapping, while ET uses the original data sample (Geurts, Ernst, & Wehenkel, 2006;Ho, 1995). AB and GBM build one tree at a time, where each new tree corrects errors made by the previous tree, an ensemble method known as boosting. AB and GBM differ in how they identify and correct errors of previous trees. AB identi es errors using high-weight data points as each new tree up-weights observations that were misclassi ed in the previous tree, whereas GBM identi es errors using gradient of loss function computed in the previous trees, which re ects how well the previous trees performed. AB corrects errors by assigning its trees different weights in the nal prediction based on their performance.
In contrast, GBM weights each tree equally, but restricts their predictive capacity through the learning rate, which represents how quickly an error is corrected from each tree to the next, for greater accuracy (J. H. We determined the baseline performance of the seven algorithms by running our data with each algorithm using thedefault hyperparameters of the algorithms in a Python 3.7 environment employing scikit-learn implementation (Pedregosa et al., 2011). We used the scikit-learn utility of Pipeline (Pedregosa et al., 2011) to automate work ow and avoid data leakage. In the work ow, we standardized the data to test all candidate algorithms. We used 80% of the data for algorithm testing and modeling and reserved 20% for validation. The training dataset and the reserved validation dataset were split randomly. We used 10-fold cross validation to estimate algorithmic performance as measured by the mean squared error (MSE).
The stochastic nature of machine learning algorithms means that model results for each run can be different. To ensure we selected the best algorithm from the candidate pool, we ran the seven candidate algorithms 100 times each on our data and averaged their performance in terms of MSE and averaged ranking out of the 100 runs.
For reference, we also used a simple multivariate parameter-based ordinary least squares (OSL) model on the data in parallel with our algorithm model as follows: where i is the sample month, REC i is residential electricity consumption for month i, Temp i is mean temperature of month i, Preci i and is the total precipitation of month i. Salary i and are salary and Pop i population levels for month i, respectively, both averaged from their yearly total. The coe cients of interest are β 0 , β 1 , β 2 , and β 3 , representing the impact of temperature, precipitation, income and population, respectively. represents the bias. We did not regroup the temperature or the precipitation data in this parallel model in order to ensure that no more assumptions were made for this model than for the algorithmic model. We tested this model with the machine learning algorithms for baseline performance. Figure 1 shows the baseline performance of the candidate algorithms using the Lanzhou and Lhasa data, respectively, with the blue bars indicating the mean MSE and the pink bars showing the averagedranking of the algorithmout of the 100 runs. It was evidentthat the ensemble algorithm GBM was the best t for studying 2011-2019 Lanzhou data and 2014-2019Lhasa data, as it showed the lowest average MSEs across the board with high consistency.

Comparing predictive accuracy
To validate the model results in case of over tting, we ran predictor variables from the validation datasets with the prepared models to make predictions. The validation results are shown in Figure 2. The good validation results suggested that the model was not over tted.

Comparing interpretability
A robust interpretation of models can only be achieved with models with high predicative accuracy. Figures 1 and 2 showed that that GBM yielded highest predicative accuracy for Lanzhou 2011-2019 data and Lhasa 2014-2019 data. To compare its interpretation with the traditional parameter-based models, we presented the results of OSL linear models across different datasets in Table 1 for reference. We used the concept of feature for discussions on PDPs and ICE plots. The concepts of "feature" and "predictor variable" are often used interchangeably, especially in machine learning models. However, we made a distinction between the two in our study for clarity. We use the term "predictor variables" to refer to the individual variables we use for modeling (i.e., temperature, precipitation, income, and population). A "feature"can either be an individual predictor variable such as temperature or two predictor variables combined into one, e.g., climate, which comprises both temperature and precipitation. To compare with the OSL interpretation based on coe cients for predictor variables, we rst interpreted the role of predictor-variables for the GBM models as shown in Figure 3. Each individual plot shows both the ICE plots and PDPs of a given feature; in this case, a predictor variable for the corresponding model, with y-axes showing partial dependence, or the marginal effect of a given feature on the REC estimation, and x-axes showing the value of the corresponding feature. Note that marginal effect is not the predicted REC value; rather, it is the way in which REC changes with the feature. It is evident that all ICE plot curves generally follow the same pattern as that shown by the PDP line. This means the PDPs of the models provide a reliable interpretation of the relationships between the features and REC.
Interpretations by GBM and OSL on individual predictor variables are generally consistent in temperature-REC relationships, with the innately non-linear GBM being more informative than OSL models. GBM suggestedthat precipitation plays a role in in uencing REC, whereas OSL appeared to differ since it failed to come to a conclusion on the role of precipitation, as shown by the high p-value and opposite signs of precipitation coe cients in the two cities in Table 1. At this stage, we cannot be sure which interpretation is correct due to the multicollinearity between year-round temperature and precipitation data, as shown in Figures 4a and 4c.
Multicollinearity may be one of the reasons why precipitation has not been fully considered in previous studies. Multicollinearity can reduce the precision of estimated coe cients in linear regressions and it is common practice to remove or lessen the weight of one predictor (in this case, precipitation). Correlations between temperature and precipitation are also a challenge for GBM interpretation, since the use of PDPs and ICE plots is based on the assumption that features are not correlated ( To test our hypothesis regarding the role of precipitation during the cold season, we created a data subset consisting of data points from November, December, January, and February, i.e., the cold-season only. Multicollinearity between temperature and precipitation in this subset was no longer a problem for either city as shown in Figures 4b and 4c. For this reason, we used this subset (rather than the year-round dataset) to analyze the individual roles of temperature and precipitation.
Quanti cation of year-round climate impact, however, needs to be conducted without interference from multicollinearity. Interpretations of an algorithm model can manage the multicollinearity between temperature and precipitation by combining them into one feature (Molnar, 2020 It is noteworthy that the PDP for the climate feature was not a simple combination of the individual PDPs for temperature and precipitation, as shown in Figure 3. Rather, the PDP for this newly created climate feature captured changes in temperature and precipitation as well as their interactions, and showed the marginal effect of all of these as a whole since partial dependence was calculated by marginalizing over all the other inputs of the estimation, i.e., income and population (Friedman and Meulman 2003; Molnar 2020).
When quantifying the marginal effect of the climate feature, we rst averaged the monthly mean temperature and monthly precipitation by calendar month to represent the corresponding mean climate patterns for the cities during their respective study periods. We then obtained possible variation ranges of the climate patterns using standard deviations of monthly mean temperature and monthly precipitation for each calendar month in the two cities. Finally, by averaging the marginal effects of all climate conditions within the variation range for each month for the two cities, we obtained the average marginal effects of climate for each calendar month in the two cities.
We tapped further into this potential of the algorithm model for holistically quantifying climate impact on REC to predict the impact of climate change on REC of the two cities.This was achieved by replacing climate observations with climate projections under RCP2.6, RCP4.5, and RCP8.5 scenarios in the PDPbased quanti cation. As our study focused on the climate impact, we effectively controlled the socioeconomic conditions at the status-quo level of the study periods. We made projections for the midcentury decade from 2036-2045, since this period is more likely to accommodate the socioeconomic presumption than more distant periods, but is su ciently distant for meaningful climate projections.

Individual roles of temperature and precipitation
The GBM and OSL interpretations described in Section 2.3.3 revealed a negative effect of temperature on REC in both cities , which is consistent with the heating "half" of the U-shaped temperature-REC function in previous studies (Auffhammer & Mansur, 2014; Li et al., 2019). Combined with local knowledge, we are con dent that the climate impact on Lanzhou and Lhasa REC is mainly driven by consumption for heating.
However, fewer references were available from previous studies on the year-round role of precipitation, and neither the OSL coe cient nor GBM PDPs were conclusive in this respect. To test our hypothesis regarding the role of precipitation in the cold season, we ran the OSL model using the cold-season data subset and obtained the results shown in Table 2. The OSL model based on the Lanzhou cold-season data subset yielded an R 2 of 95.9%, and considered precipitation to be a statistically signi cant predictor variable with a p-value lower than that of temperature. The positive coe cient of precipitation suggested that more precipitation in the cold season will drive up REC. The cold-season subset was too small to model using GBM; however, we were able to obtain con rmation based on its year-round interpretation. In Figure 4a, GBM shows that the marginal effects of precipitation on REC are higher when monthly precipitation falls below 40 mm in Lanzhou and below 20 mm in Lhasa. Climate patterns of the two cities shown in Figures 5c and 5d indicate that such low monthly precipitation only occurs in the cold season, suggesting that the impact of precipitation is stronger during the cold season.
The OSL model for the Lhasa cold season yielded an R 2 of 88.4%, and suggesteda negative precipitation-REC relationship. There was no statistically signi cant precipitation-REC relationship in this case, though the p-value for precipitation was signi cantly lower for the Lhasa cold-season subset than for the Lhasa year-round data. Despite this, we consider that the result can be used as a reference, since it would be inappropriate to conclude that there is no association simply because the p-value is > 0.05 (Amrhein, Greenland, & McShane, 2019). The GBM interpretation of the precipitation impact in Lhasa shown in Figure 4b also con rmed this theory, as it showed a small peak in a low precipitation range in cold season, suggesting the existence of a precipitation impact during the cold season.
We identi ed an explanation of the absence of a positive OSL coe cient for precipitation in Lhasa cold season by comparing the Lhasa climate pattern with that of Lanzhou, as shown in Figure5b. The monthly mean temperature rarely drops far below 0 °C in Lhasa, even in the cold season. Additionally, monthly precipitation during the cold season in Lhasa is considerably lower than that duringthe cold season in Lanzhou. In other words, compared with Lanzhou, snowfallis not as frequent or as heavy in Lhasa; hence, the less observable impact of precipitation during the cold season in Lhasa.
We conclude that the impact of precipitation is stronger in both cities during the cold season, as a result of snowfall boosting electricity consumption used for heating.

Impact of climate
As discussed in section 2.3.3, quantifying marginal effects of year-round temperature and precipitation separately would not be accurate due to the presence of multicollinearity. It would beunreasonabletoomit either temperature of precipitationfrom our estimation just because of their numerical correlation, since our studyalready proved that both play a role as independent climate variables in in uencing REC. For this reason, we quanti ed their joint impact (i.e., the climate impact) by interpreting the climate feature enabled by PDP. Figures 5a and 5c show the marginal effects of climate on REC, and the climate impact on REC in Lanzhou and Lhasa, respectively. We used real-world climate patterns in Figures 5b and 5d to assist in constrainingthe way in which we used PDPs so that conclusions were not made based on extremely unlikely combinations of temperature and precipitation. The calculated average marginal effect showing how REC changes with climate conditions is marked in purple in Figures 5b and 5d.
Note that the z-axes of the PDPs in Figures 5a and 5c show the marginal effect of climate on REC rather than the actual predicted REC values; this describesthe means by and degree to which REC changes with climate, i.e., the climate impact. The same is true for the purple y-axes in Figures 5b and 5d. A climate impact of zero showed that the climate feature did not have anyeffecton the estimation. Positive climate impact values show that the corresponding climate conditions will drive up the estimated REC and that the negative values will reduce it. For ease of reference, we describe changes in the climate impact by referring to them as positive and negative climate impacts; i.e., when we state that the climate impact is enhanced, it means it has increased in absolute value.
Interpretation of the GBM model using a holistic climate feature in Figure 5b shows a V-shaped relationship between climate impact on REC and calendar months for Lanzhou, with July showing the strongest negative impact from climate and November, December, January, and February showing the greatest positive climate impact. The climatic impact change during cold-season is moreconsistent with the change in precipitation, which consolidated our previous nding hat precipitation has stronger impact on Lanzhou REC during cold season, likely in the form of snowfall given the mean temperature of the cold season months areslightly above or below 0°C.
For Lhasa, the relationship is U-shaped (Fig.5d), with June, July, August, and September showing similarly strong negative climate impacts and December and January showing the greatest positive climate impacts. The climatic impact change during cold-season is more consistent with the change in temperature than in precipitation. This can be explained by less snowfall in Lhasa. Precipitation is low in Lhasa during cold season where there is only one month with mean temperature below 0°C.
Given the climate patterns, this indicates that climatic impact on REC is mainly driven by heating demand in the two cities, and that Lanzhou REC is more sensitive to climate variations in warm months than Lhasa.
We also noticed that there is not a consistent temperature threshold for climatic impact on REC in the two cities, as opposed to the assumption of HDD and CDD that there is a generally consistent base temperature at which there is no energy demand for cooling or heating. Climatic impact on REC is around zero when monthly mean temperature is around 12°C for Lanzhou and around 8°C for Lhasa. This further proves our point that temperature alone does not determine the climatic impact on REC. We believe precipitation as well as socioeconomic factors may also play a role in determining the climate threshold where there is no signi cant energy demand for cooling or heating.

Projected change in climate impact on REC
For the meaningful comparison of climate impact on REC under status-quo and projected RCP2.6, RCP4.5, and RCP8.5 scenarios, insights are needed regarding bias of the simulated climate patterns; the smaller the simulated bias, the more reliable projected climate impact. Lanzhou. This means precipitation has an offsetting effect to temperature on cold-season REC in Lanzhou.
For the remainder of the year, only July, August, and September showed enhanced negative climate impact. This also suggested an offsetting role of precipitation since warming alone would only enhance the negative climate impact. Since July and August (i.e., the two warmest months in Lanzhou)will still experience a negative climate impact, we do not believe that the drop in negative impact in other months will be driven by cooling-related demands. Rather, a more logical explanation would be that more precipitation may have a cooling effect which will offset the warming impact on REC. This offsetting impact of precipitation may be adequately strong to change the climate impact. For example, in October, under RCP 2.6, the climate impact changed from negative to positive, suggesting that the climate impact changed from reducing the REC estimation to increasing it. This can be con rmed by the fact that the negative climate impact was generally more enhanced under RCP 8.5 than under RCP 2.6.The varied ranking of climate impact under RCP4.5 among the three scenarios suggests that under this scenario, the effects of precipitation and temperature have competing importance in shaping the overall climate impact.
We did not thoroughlyinterpret the quanti cation of the climate impact on REC inLhasa as the simulated climate was considered to be excessively biased to elucidate the projected climate impact on REC.
However, as Lhasa,like Lanzhou, is becoming warmer and wetter, we also infered precipitation to offset the warming impact on REC. However, as Lanzhou showsstronger precipitation impact in the cold season and is more sensitive to climate in warm months, we expect Lanzhou to be affected more than Lhasa by this offsetting effect.

Discussion And Conclusions
Our study proves the strength of algorithm modeling since it shows improved predictive accuracy and interpretability for studying climate impact on REC than linear models. Compared with previous econometric models used for climate-REC studies, our method allows highly correlated temperature and precipitation to be interpreted as one climate feature to quantify the climate impact. This means that our method has overcome the challenges of multicollinearity, which enables our models to treat temperature and precipitation equally, and has the potential to be used in all climate impact studies.
A data-driven model can only be as good as the data. Our model failed to make predictions for several scenarios that exceeded the bounds of the PDPs of climate feature. Our study was limited as we controlled the non-climate inputs at the status-quo level to predict climatic impact on REC; however, socioeconomic conditions will de nitely alter REC and, thus, the marginal effects of climate impact. Future studies should have a socio-economic focus to enable more accurate predictions. Our projection for Lhasa is also limited by insu cient climate projection studies in the region. More accurate climate projections are needed to improve the understanding of potential future climate impact on REC.
Our results show that the climate impact on REC in both Lanzhou and Lhasa is driven by heating-related demand, with Lanzhou being more sensitive to climate than Lhasa during the warm season. The model results support our hypothesis. Both Lanzhou and Lhasa show a stronger precipitation impact during the cold season, since precipitation-especially in the form of snowfall-boosts electricity consumption during the cold season. The impact is stronger in Lanzhou, which experiences greater precipitation and more months when temperatures below 0 °C occur during the cold season. As the two cities become warmer and wetter, precipitation will offset the impact of warming on REC across the year, with Lanzhou being more affected. Based on Lanzhou projections, the offsetting effect is more pronounced in cold season across all scenarios, and is particularly strong under RCP 2.6. For the remainder of the year, the effects of increased precipitation and warming will have competing importance under the RCP 4.5 scenario, while the effects of temperature will generally dominate the climate impact under the RCP 8.5 scenario.
Our study highlights the need for climate-related energy studies to address climate holistically rather than to focus solely on the temperature, especially in regions with cold winters as well as in areas vulnerable to extreme weather such as blizzards. It provides previously unavailable references for future climateenergy studies in western China and broader alpine regions where temperature is rapidly increasing just as the water cycle is becoming ever more unstable. Our study also assists in aiding regional policymaking on climate change adaptation. The developed model is also readily compatible with potential early-warning projects to protect against power outages induced by extreme weather.

Declarations
Data availability The data that support the ndings of this study are available from data sources as stated in section 2.2.
Restrictions apply to the availability of these data, which were used under license for this study. Data are available from the authors with the permission of data sources.