Short-term load forecasting using neural networks and global climate models: An application to a large-scale electrical power system

This paper focuses on the development of shallow and deep neural networks in the form of multi-layer perceptron, long-short term memory, and gated recurrent unit to model the short-term load forecasting problem. Different model architectures are tested, and global climate model information is used as input to generate more accurate forecasts. A real study case is presented for the Brazilian interconnected power system and the results generated are compared with the forecasts from the Brazilian Independent System Operator model. In general terms, results show that the bidirectional versions of long-short term memory and gated recurrent unit produce better and more reliable predictions than the other models. From the obtained results, the recurrent neural networks reach Nash-Sutcliffe values up to 0.98, and mean absolute percentile error values of 1.18%, superior than the results obtained by the Independent System Operator models (0.94 and 2.01% respectively). The better performance of the neural network models is confirmed under the Diebold-Mariano pairwise comparison test.

Automated summary

This paper investigates the application of shallow and deep neural networks in modeling short-term load forecasting problem. Different model architectures including multi-layer perceptron, long-short term memory, and gated recurrent unit are tested, and global climate model information is used as input for more accurate predictions. A case study for the Brazilian interconnected power system is presented and compared with forecasts from the Brazilian Independent System Operator model. The results show that bidirectional long-short term memory and gated recurrent unit outperform other models, achieving Nash-Sutcliffe values up to 0.98 and mean absolute percentile error values of 1.18%, superior to the results obtained by the Independent System Operator models (0.94 and 2.01% respectively). The better performance of neural network models is confirmed under the Diebold-Mariano pairwise comparison test.