Auto-encoder Neural Network-Based Monthly Electricity Consumption Forecasting Method Using Hourly Data

The effectiveness of monthly electricity consumption forecasting (ECF) directly affects the profitability of electricity retailers in a deregulated market. The monthly ECF using fine-grained hourly data based on the multi-step forecasting strategy normally shows unsatisfactory performance, given the fact that it contains numerous forecasting steps. Aggregating the data points is a common approach which can reduce the forecasting steps by compressing the data series. However, the information loss caused by the additive aggregation method generally leads to low predictability of the compressed data series. To address this challenge, we propose an auto-encoder neural network (AENN) based data compression method. Specifically, an AENN with a small central layer is first trained to reconstruct the fine-grained hourly electricity consumption input data. Subsequently, the former part of the trained AENN is used to compress the hourly data into the coding series. Then, the multistep forecasting model is trained based on the coding series. Finally, the forecast result of the coding is decoded using the latter part of the trained AENN to form the electricity consumption forecast. Numerical experiments demonstrate the superiority of the proposed method while combined with three representative AI forecasting algorithms.