Data driven approach for short-term electrical load forecasting

Abstract

This thesis work aims to develop an evolutionary deep learning-based hybrid data-driven approach for short-term load forecasting (STLF) in the context of Bangladesh. With the lapse of time, the power system is getting complex. Penetration of intermittent renewable energy (RE) into the grid, changing prosumer load patterns with the need for demand side management (DSM) has thrown a challenge for dynamic power system operation and control. Load forecasting plays a significant role in this dynamic operation and control. In addition, it directly affects the future planning of network expansion, unit commitment, and economic energy mix for the power market. The day ahead short-term forecasting is very crucial in day-to-day operation. As such, various conventional and modified methods have been used over time for a short-term prediction. Nevertheless, the existing approaches like age-old statistical methods, artificial intelligence (AI), machine learning (ML), and deep learning (DL) techniques alone cannot provide effective accuracy all the time. Hence, an integrated genetic algorithm (GA)-bidirectional gated recurrent unit (BiGRU) hybrid data-driven technique (GA-BiGRU) is proposed in this work. The developed method is validated in the Bangladesh power system (BPS) network by providing day ahead forecasting of the electrical load of the whole country. Besides, the performance of the prediction model is compared with some existing approaches such as long short-term memory network (LSTM), gated recurrent unit (GRU), and integrated genetic algorithm-gated recurrent unit (GA-GRU) in terms of mean absolute performance error (MAPE) and root mean squared error (RMSE) as well as coefficient of determination (R2). The outcome gives an indication of better forecasting accuracy of the proposed GA-BiGRU evolutionary DL technique compared to others. It has been found that GA optimized BiGRU model with two hidden layers provides a minimum 18.13% and 19.82% reduction in RMSE and MAPE respectively for day ahead prediction while RMSE and MAPE reduction of 16.01% and 16.39% respectively for the case of 48 hours ahead projection.