Master's Thesis : Hyperparameters and features selection for forecasting energy generation and consumption

Abstract

Microgrids are electrical power systems made of consumption, generation, and storage devices. Their emergence raises some questions concerning the generation of reliable energy forecasts. Indeed, each microgrid has its characteristics, it can integrate different energy resources and be subjected to different loads. Moreover, the production of these energy resources and the load can be subjected to many variations over time (due to the weather, the inherent characteristics of the resource, socio-economic factors). Therefore, the automatic generation of energy forecasts is a challenging task. This thesis develops a methodology for selecting the hyperparameters and features of existing forecasting models and for using those optimized models for autonomously generating forecasts.

To this end, several forecasting models were optimized. These models were statistical time series models such as the naive forecaster or the exponential smoothing technique and some artificial intelligence-based models such as the linear regression, the gradient boosting, or the multilayer perceptron. Since the performances of different models can vary over time, the usage of a multi-model forecasting system is then analyzed. Multi-model forecasting enables to periodically calibrate the weights assigned to the different models and thus improves the final forecast. The multi-model methods tested are based on the linear combination of the different models. Firstly, some simple multi-models producing the mean or the median of the forecasts. Then, some methods based on inverse MSE weighting and linear regression weighting have been tested.

The individual models that produced the best results are the gradient boosting regression and the multilayer perceptron. The multi-model forecasting method that led to the best results is the multi-model selecting the median of the forecasts. It is also the method with the lowest costs in terms of computational time.