Enhancing the prediction of electric load demand: a comparative analysis of ARIMA and ANN models for the case of a small touristic island

Accurate prediction of energy demand is necessary for efficient power system operation, particularly in energy systems with high renewable sources integration. This study compares different methods for predicting load demand using statistical regression. In particular, the goal is to provide insights on the differences between simpler AutoRegressive Integrated Moving Average (ARIMA) and more complex Artificial Neural Network (ANN) models. The algorithms are used to predict electric load demand in Tilos, a small Greek island with strong seasonal trends due to summer tourism. The case study is significant as Tilos' outdated electrical grid must adapt to an increasing share of renewable energy sources, making load forecasting increasingly important. The algorithms were developed in Python using open-source tools (such as StatsModels and TensorFlow). Hyperparameters’ tuning, crucial for enhancing forecasting effectiveness, was performed using stochastic optimization with Differential Evolution to minimize RMSE. The optimal normalized RMSE was reported as 9.72% for ANN and 9.54% for ARIMA, showing the effectiveness of both methods, with a slight edge for the statistical model. This work provides critical information regarding load prediction methodologies, highlighting practical guidance for energy managers, policymakers, and researchers in power system planning and operation.