Electric energy demand forecasting: the case of Aydin, Denizli, and Muğla regions

This study aims to forecast the hourly electricity consumption of Aydın, Denizli, and Muğla provinces by addressing the challenges faced by distribution and retail companies operating in the Turkish electricity market during the development of energy prediction algorithms. The primary objective is to enhance forecast accuracy in energy distribution, analyze seasonal and regional demand variability, and identify which models yield the most successful results. Accordingly, hourly electricity demand forecasts were conducted using CatBoost, BiLSTM, and Transformer models. Electricity consumption forecasts were made on a 24-hour basis across two different scenarios and timeframes. The purpose of these scenarios is to simulate the outcomes of Day-Ahead Market (DAM) and Intraday Market (IDM) operations. In the first scenario, the forecast for each day was made using data up to the last hour of the previous day (t-1). In the second scenario, data from two days prior (t-25) were used. Additionally, 15 day forecasts were made for March and July to observe how the models responded to different time periods. This approach allowed for a detailed performance comparison of the models under varying conditions. Forecast performance was evaluated using the MAPE (Mean Absolute Percentage Error) and RMSE (Root Mean Squared Error) metrics. The results demonstrated that the CatBoost model consistently delivered the most accurate predictions across both scenarios. While the Transformer model produced results comparable to CatBoost during March's stable demand patterns, its performance declined under the volatile demand patterns observed in July. On the other hand, the BiLSTM model exhibited higher error rates in both scenarios due to its limited ability to capture short-term fluctuations in time series data. The findings indicate that the approaches used in this study enable reliable forecasts for both Day-Ahead Market (DAM) and Intraday Market (IDM) operations. Furthermore, these results suggest that distribution and retail companies can develop strategies to minimize financial losses in market transactions by leveraging these predictive models effectively. This study emphasizes the necessity of adapting electricity demand forecasting models to seasonal and regional dynamics, offering significant contributions to the development of data driven strategies in energy management processes. The findings provide valuable insights for improving the reliability and efficiency of energy planning and market strategies, serving as a practical guide for stakeholders in the energy sector.