Empirical comparison of load forecasting methods for Skagerak energilab : a perspective of the operational and economic efficiency gain as a result of increased forecasting accuracy in a microgrid environment

Abstract

This master thesis is analyzing short-term load forecasting. Power consumption in kW will be forecasted 24 hours ahead, for each day of a week and finally averaged to derive mean performance. The forecast will be conducted by selected methods and models and compared against a simple yet reasonable benchmark model. To evaluate the performance in detail, we select to compute MAPE values for each individual hour, day and average over one week. In addition, we construct a tailored evaluation metric to estimate the economic consequences of inaccurate load forecasts. This master thesis is intended to provide a theoretical and empirical link between contemporary forecasting techniques and actual economic benefits that can be derived from improved accuracy of load forecasts at Skagerak Energilab. Obtained results show a tendency of increased forecasting accuracy when utilizing machine learning algorithms with Neural Network structures. However, no single method could outperform an ensemble average model. Compared to the benchmark model, our proposed Ensemble consisting of BATS, seasonal ARIMA, and a multivariate AR ANN increased forecasting accuracy by a notable degree. Also, improved performance was shown to result in a decreased direct economic cost.