Next-day Load Forecasting with Smart Meter Data Using Global Recurrent Neural Networks

Abstract

In a smart grid, consistent and accurate load forecasting is critical to successful operation and energy management. With the growing availability of smart-meter data, machine learning models have the ability to train one model globally across entire datasets, rather than training a separate model individually for each time series. A global training set-up enables the learning of patterns of electricity consumption in households and brings potential computational advantages. We present an experiment that evaluates a Long-Short-Term Memory (LSTM) Global Forecasting Model (GFM) in comparison to benchmark methods: Feed-Forward Artificial Neural Network, Seasonal Autoregressive Integrated Moving Average and standard load profile models. The selected methods and the evaluation framework are derived from an extensive literature review. We use the Irish smart-meter dataset, collected by the Commission for Energy Regulation. Assessment includes the scalability and computational performance of the algorithms. Comparisons show that the average error in terms of several metrics is at least $3\%$ smaller than the benchmark performance, indicating that the LSTM-GFM model obtains predictions with a superior accuracy. A complementary 'weak learner' model, used to generate features from a seasonal decomposition, further decreases the error. The study shows that LightGBM models are faster and more suitable for quick model iterations and LSTM-based models are more appropriate for accuracy-focused load forecasting applications.