Probabilistic Wind Power and Electricity Load Forecasting with Echo State Networks

With the introduction of distributed generation and the establishment of smart grids, several new challenges in energy analytics arose. These challenges can be solved with a specific type of recurrent neural networks called echo state networks, which can handle the combination of both weather and power consumption or production depending on the dataset to make predictions. Echo state networks are particularly suitable for time series forecasting tasks. Having accurate energy forecasts is paramount to assure grid operation and power provision remains reliable during peak hours when the consumption is high. The majority of load forecasting algorithms do not produce prediction intervals with coverage guarantees but rather produce simple point estimates. Information about uncer- tainty and prediction intervals is rarely useless. It helps grid operators change strategies for configuring the grid from conservative to risk-based ones and assess the reliability of operations. A popular way of producing prediction intervals in regression tasks is by applying Bayesian regression as the regression algorithm. As Bayesian regression is done by sampling, it nat- urally lends itself to generating intervals. However, Bayesian regression is not guaranteed to satisfy the designed coverage level for finite samples. This thesis aims to modify the traditional echo state network model to produce marginally valid and calibrated prediction intervals. This is done by replacing the standard linear regression method with Bayesian linear regression while simultaneously reducing the di- mensions to speed up the computation times. Afterward, a novel calibration technique for time series forecasting is applied in order to obtain said valid prediction intervals. The experiments are conducted using three different time series, two of them being a time series of electricity load. One is univariate, and the other is bivariate. The third time series is a wind power production time series. The proposed method showed promising results for all three datasets while significantly reducing computation times in the sampling ste