Forecasting Electricity Consumption In Kyrgyzstan Using Machine Learning

Power demand is a barometer of economic growth and industrial activity. Electricity is essential for domestic heating, agricultural irrigation and industrial processes in Kyrgyzstan, a Central Asian nation that is densely mountainous and rich in hydro. With the country striving to achieve greater energy sustainability and to satisfy growing future demand, the importance of accurate forecasting cannot be overstated. References ARIMA and other traditional forecasting methods have been traditionally used, however in today’s landscape, ML techniques hold promise in terms of accuracy and flexibility for forecasting.

The National Statistical Committee of the Kyrgyz Republic (2023) also reports that the electricity consumption per capita has been increasing steadily, especially in urban areas where the process of modernization is also more pronounced. It is also reported in the Energy Ministry (2022) that the demand for peak load during winter season has been increasing, making load forecasting a crucial part of the management plan.

The goal of this study is to test and examine different ML models for Kyrgyzstan electricity consumption forecasts based on history data from the World Bank. We compare results of ARIMA model to those produced by the Random Forest, XGBoost and CatBoost algorithms, focusing on MAE, RMSE and R-squared metrics.

Hydropower dominates Kyrgyzstan's power sector, providing around 90% of its electricity supply. This would make the country one of the greenest energy producers in Central Asia, while at the same the time making it highly vulnerable. The seasonal ebb and flow of water levels, due to snowmelt, drought or climate change, are major sources of unpredictability in power supply, particularly in winter when demand climbs but hydro production falls.

In addition, the electrification of the economy, extension of the cities and digitalization increase electricity consumption. Nevertheless the current predictive modelling used in national planning does not take easily to nonlinear trends in consumption, or how these are influenced by extraneous elements e.g. hydrological cycles and industrial development. This discrepancy in prediction would lead to supply-demand imbalances, blackouts, and suboptimal utilization of energy resources.

The lack of reliable, data-based forecasting is equally detrimental to the defense and energy objectives, in particular on investments in grid infrastructure, imports of winter energy and de-risking cross-border energy markets. Hence, there is the pressing need for better forecasting tools that can reliably predict electricity consumption, take into account the supply-side volatility from hydropower, and inform energy policy in Kyrgyzstan.

The dataset used in this study is sourced from the World Bank, specifically the indicator "EG.USE.ELEC.KH.PC" representing electric power consumption (kWh per capita). The time series spans from 1997 to 2022.

Missing value handling using linear interpolation.

Log transformation to stabilize variance.

Normalization for use in ML models.

This thesis showed the great potential of machine learning models in predicting electricity consumption for Kyrgyzstan — a country characterized by the seasonality of hydro-electricity generation and an increasing residential consumption. By utilising historical consumption data sourced from the World Bank, as well as advanced forecasting methods, this research has provided valuable insights into the relative performance of traditional statistical-based and contemporary machine learning forecasting techniques. Of the tested models, the Random Forest method obtained the most stable and accurate prediction, with an R² value as high as 0.9526. This demonstrates that the pattern hidden in the electricity consumption data in Kyrgyzstan is highly complicated and nonlinear, and ensemble learning methods are very suitable to capture such patterns. Finally, CatBoost and XGBoost also achieved good performance, whereas ARIMA, being an interpretable technique frequently used for classical time series forecasting, turned out to be less accurate and not so well adapted to the recent consumption trends. These findings are especially significant in the face of heavy dependence on hydropower in Kyrgyzstan. The nation encounters major issues of seasonal water flows, unpredictable environmental conditions, and high energy consumption in the winter period. The poor ability of conventional methodologies to deal with these factors with a high degree of accuracy argues for the necessity of including increasingly dynamic, information- driven tools in national energy planning. This research adds to the body of existing work by being one of the earliest to consider the Kyrgyz energy system in particular by focusing on long-term load forecasting based on the machine learning models. It fills a significant research void, and it offers methodological innovation and a useful toolkit which could be made use of by energy planners, policymakers, and infrastructure investors.

Besides demonstrating the relevance of machine learning the thesis emphasizes the necessity for reliable and transparent data sources. The using of WB data and the NSC of KR is also a significant basis for predictability and replicability of study. On the other hand, the study isn't without limitations, the researchers admit. Although the models presented here worked well, they were essentially constructed from univariate time series information. In future, multivariate characteristics (like changes in temperature, population growth, industrial production, energy prices, political movement, and so forth) of the data should be incorporated to obtain better generalisation and interpretation of models. This would further enhance the utility of the models for online applications and decision support systems. Further, to deploy such forecasting models in practice, such as in the Ministry of Energy or in utility companies, would necessitate user-friendly notebooks and pipelines for continuous learning. The scope of the long-haul vision may embrace the deployment of these models in combination with that of smart grid, so as to achieve an automatic response to forecast load fluctuations, especially in seasonal peaks. Conclusively, this study demonstrates that machine learning improves the accuracy and reliability of short-term electricity consumption forecasts for Kyrgyzstan. By moving away from conventional methods to smarter, adaptable systems, the nation can gain in terms of energy security, resources efficiency and policy effectiveness. In the face of increasing energy demand, the use of such tools is essential to ensure the sustainability of development and resilient energy future for Kyrgyzstan.