Load Forecasting with Deep Learning

Abstract

Load forecasting has emerged as one of the most significant subjects of study in recent years. The load curve is the demand curve of the microeconomic paradigm determining the prices in power markets, the other curve is that of the supply. In our effort to contribute to the field, we apply several Deep Learning techniques and then compare their results. Before that, we used simple data like lagged load values and temperatures. In addition, we account for the hour, day, and season effects with simple data handling. Our results are accurate, and our models can generalize well. We find that simple models like LSTM Encoder-Decoder and MLP are better at time-series forecasting than more complex models like Transformers. The simplicity of our models and data presents evidence that researchers should not go to great lengths for accurate forecasting. A balance should be retained between data availability (ours are free), computation expenses, results, understandability, and accuracy. We believe that our research has achieved a good level of balance in those fields.