Fitting multiple temporal usage patterns in day-ahead hourly building load forecasting under patch learning framework

This paper proposes a novel day-ahead hourly building load forecasting approach under the framework of patch learning, a recently proposed data-driven model that aggregates a global model and several patch models to further reduce forecasting errors. A patch learning model based on the long short-term memory network is hereby employed to address such a time-series-based forecasting problem, where the long short-term memory network is considered as the global model and the support vector regression is selected as the patch model. To obtain satisfying performance, the largest absolute error measurement is selected to evaluate load forecasting errors and identify patch locations. Furthermore, a genetic algorithm with an elitist preservation strategy and the grid search method are employed for hyperparameter tuning of the global model and patch models, respectively. The performance of the proposed model is tested and verified on two practical building load data sets and the Lorenz chaotic time-series data and compared with four advanced building load forecasting models on several common metrics.

Automated summary

This paper proposes a novel day-ahead hourly building load forecasting approach using the patch learning framework, which combines a global model and patch models to reduce forecasting errors. The performance of the proposed model is evaluated on practical building load datasets and compared with four advanced building load forecasting models on common metrics.