LSTM-based Active and Reactive Load Forecasting and its Replicability in Large Geographical Areas
The day-ahead load forecast is essential for the efficient planning and operation of electric power systems, especially in the context of smart grids. This task is becoming increasingly important with the growing integration of variable renewable energy sources. Among the various machine learning-based load forecasting methods, Long Short-Term Memory (LSTM) networks have shown to be particularly effective. This paper analyses the impact of reactive load as an exogenous variable on active load forecasting and vice versa, employing LSTM networks with hyperparameters optimized through Gaussian Process Regression (GPR). The results, validated using dataset from Bangalore, India, demonstrate that including exogenous variables enhances forecasting accuracy. Additionally, the effect of different training/(validation+test) percentage ratios on prediction performance is evaluated finding that a 70%:30% ratio yields a satisfactory balance of accuracy and training efficiency. Finally, a combined forecasting model is used to analyse the forecasting accuracy of a model that is trained using data from one location (Bangalore) and tested using data from another location (Itanagar), proving there is no overfitting in the forecasting model.