Comparative Analysis of Techniques for Short-Term Electrical Load Forecasting

This study addresses the challenge of accurately forecasting electricity load in Pakistan, focusing on the Faisalabad Electric Supply Company (FESCO). The load forecasting problem in this region is exacerbated by the highly volatile nature of the data and the low baseload, further complicated by external factors such as weather conditions. To tackle this issue, we utilized historical electricity load data from FESCO from 2019 to 2022 and weather data from NASA’s LaRC POWER Project. Our approach involved comprehensive exploratory data analysis (EDA) to identify significant input features, including temperature, humidity, and lagged predictors like previous hour and previous day readings. We employed a range of deep learning models to develop and test prediction models like long short-term memory (LSTM), bidirectional LSTM (BiLSTM), and gated recurrent unit (GRU) networks. The analysis revealed that lagged predictors significantly enhance prediction accuracy, with BiLSTM models demonstrating the best performance, achieving a remarkably low mean absolute percentage error (MAPE) of 0.2%. Compared to other models, our approach using time-series data arrangement without external weather predictors proved to be more accurate and economical. This model can support effective power system planning and expansion, leading to the development of a competitive bidding-based wholesale energy market in Pakistan.