AI-Based Predictive Analysis for Energy Demand Forecasting in Pakistan: Incorporating Carbon Credit Indicators

Abstract

Energy demand forecasting is highly important in the energy sector from generation planning, fuel procurement, and tariff calculation point of view in Pakistan due to the high economic cost of mismatch between energy supply and demand. Non-linear patterns are hard to predict using traditional forecasting approaches and climate-policy signals are not explicitly included in these approaches, which are becoming more important in energy markets. This research compares the accuracy of predictive models based on artificial intelligence (AI) techniques namely gradient boosting (XGBoost), and long short-term memory (LSTM) networks and assesses whether they enhance the accuracy of electricity demand forecasts in Pakistan, and whether carbon credit indicators, such as carbon prices from EU Emissions Trading System (ETS) provide value as an indicator of electricity demand. Secondary monthly time series data from January 2015 to December 2024 was used and a quantitative, longitudinal research design was adopted. Electricity demand and consumers data was gathered from NEPRA, energy prices from OGRA and PPIS, carbon prices from EU ETS, and temperature from PMD. The modelling framework included benchmark models (multiple regression, SARIMAX) and AI models, with the models validated using time‑ordered out‑of‑sample tests. The performance was evaluated using MAE, RMSE and MAPE. The Boost model significantly outperforms SARIMAX model, with a MAE of 290 MW compared to 380 MW for SARIMAX. The second most important predictor was economic activity which supported its positive association with demand (H1). Adding carbon price features increases the accuracy of the forecasts by 13.4% (p = 0.03) and is among the top five most important features, thereby supporting H2. The importance of temperature did not change (33.5%). The results of LSTM were comparable but not significantly better than Boost. The study findings conclude that the use of AI ‐forecasting techniques, especially gradient boosting, greatly improves the accuracy of energy demand forecasting in Pakistan. Even in the absence of a domestic carbon market, it is important that carbon credit indicators include relevant predictive information, and that they should be included in energy planning scenarios. The recommendations are to switch to Boost for operational forecasting, to track international carbon prices, and to enhance the energy data infrastructure in the country.