Data-driven Approaches Predict Hourly Electricity Demand Profiles at Industry and City-level

Abstract

Accurate electricity demand forecasting at industry and city-level is vital for precise demand-side management. This paper proposes an hourly electricity demand forecasting algorithm at industry-level and city-level based on load decoupling and a prior-knowledge-adjusted forecasting model. The proposed method decouples electricity demand firstly into basic electricity demand and the temperature-induced demand, in which the temperature-induced demand is predicted by a data-driven deep learning model (Bi-directional Long-Short-Term Model, BiLSTM) to achieve hourly-level forecast. The BiLSTM model can exploit the bidirectional relationship of time series data to deeply mine feature information. Moreover, prior knowledge is applied to the model parameter optimization, which solves the calendar effect without using calendar features like holiday flags or week indexes. Historical electricity load, temperature, heat index, and comfort index are considered in the forecast model. The 2021 hourly electricity load data of industries in Fuzhou and Xiamen city, China is used as benchmark data for evaluation. The performance of the proposed algorithm is further compared with LSTM, GRU, and RNN algorithms indicating that the proposed algorithm is 41.7% more accurate than the LSTM, the prediction error of city-level electricity demand is 1.09%, and the forecast errors of industries are almost no more than 2%.