Machine learning long-term electricity demand forecasting system for strategic energy investments

Abstract

The Energy Institute’s 2025 report revealed that the global electricity consumption is rising by 4% annually, with regional variances driven by complex economic, social and climatic factors. Given the critical role of long-term consumption forecasting in informing energy investment decisions and ensuring the resilience of national power infrastructure against future demand fluctuation risks, the development of demand prediction systems that are both accurate and robust is imperative. Current univariate—assuming stable demand patterns, and multivariate models—constrained by narrow feature sets, possess a limited ability to capture cross-domain interactions among economic, social and climatic factors. To address these gaps, this study introduces a forecasting system that encompasses three-phases: i) Data Acquisition and Structuring; ii) Predictive Modeling and Optimization; and iii) Model Evaluation and Analysis. The developed system integrates multi-domain predictors, including electricity generation, as well as economic, social and climatic features, while also leveraging deep learning and SHAP-based interpretability to ensure accurate and robust predictions and provide insights into feature contributions. Deployed on Egypt using data from 2000 to 2023, the system results indicate that population and GDP per capita are the primary drivers of electricity demands, with generation capacity and external debt exerting secondary influences. The system model achieved a mean R²=0.83 across multiple random weight initializations and a root mean square error of 4.1% of the mean testing value, demonstrating its reliability for long-term scenario-based planning, strategic investment, and energy security policy formulation.