Machine Learning methods for long and short term energy demand forecasting

The thesis addresses the problems of long- and short- term electric load demand forecasting by using a mixed approach consisting of statistics and machine learning algorithms. The modelling of the multi-seasonal component of the Italian electric load is investigated by spectral analysis combined with machine learning. In particular, a frequency-domain version of the LASSO is developed in order to enforce sparsity in the parameter and efficiently obtain the main harmonics of the multi-seasonal term. The corresponding model yields one-year ahead forecasts whose Mean Absolute Percentage Error (MAPE) has the same order of magnitude of the one-day ahead predictor currently used by the Italian Transmission System Operator. Again for the Italian case, two whole-day ahead predictors are designed. The former applies to normal days while the latter is specifically designed for the Easter week. Concerning normal days, a predictor is built that relies exclusively on the loads recorded in the previous days, without resorting to exogenous data such as weather forecasts. This approach is viable in view of the highly correlated nature of the demand series, provided that suitable regularization-based strategies are applied in order to reduce the degrees of freedom and hence the parameters variance. The obtained forecasts improve significantly on the Terna benchmark predictor. The Easter week predictor is based on a Gaussian process model, whose kernel, differently from standard choices, is statistically designed from historical data. Again, even without using temperatures, a definite improvement is achieved over the Terna predictions. In the last chapter of the thesis, aggregation and enhancement techniques are introduced in order to suitably combine the prediction of different experts. The results, obtained on German national load data, show that, even in the case of missing experts, the proposed strategies yield to more accurate and robust predictions.