From Data to Demand: Advanced Electricity Load Forecasting Using Neural Network

Accurate electricity demand forecasting is essential for reliable grid management, especially in energy systems undergoing rapid integration of renewable sources. This thesis presents a deep learning approach based on Long Short-Term Memory (LSTM) networks for short-term electricity load forecasting in Italy using high-resolution time series data from 2022 to 2024. The proposed model was trained on 15-minute interval load data and evaluated using standard metrics, including Mean Squared Error (MSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error(MAPE) and Relative Mean Absolute Error(RMAE). To benchmark its effectiveness, the model’s forecasts were compared against those provided by Terna, the national transmission operator. Comparative analysis revealed that the developed LSTM model yielded higher prediction errors than Terna’s official forecasts, with the discrepancy potentially attributed to the high resolution forecasting interval (15 minutes versus hourly intervals commonly used in similar studies) and the absence of supplementary factors such as weather conditions and economic indicators that significantly influence load variability. These findings highlight the potential of deep learning methods to enhance the accuracy of operational load forecasting, offering a scalable and adaptive alternative to conventional approaches in real-world power systems.