Separate wind power and ramp predictions based on meteorological variables and clustering method

Much recent research in wind power forecasting has been focused on predicting large, sudden changes in wind power output, called wind ramps. However, the current wind power forecasting methodology, which combines Numerical Weather Prediction (NWP) models and machine learning methods, has limitations in addressing the prediction accuracy in different weather conditions. Based on existing wind power forecasting methods, this paper proposes a separate power forecasting method that addresses different weather regimes. A framework for wind power predictions is first built, which uses wind forecasts from a Weather Research and Forecasting Model (WRF) model as input and converts the input into future wind power generation using a Multi-Layer Perceptron Neural Network (MLP-NN). Specific power prediction systems are then built for each subset of data, which is divided according to the hourly wind speed changes, the synoptic atmospheric circulation types, and the K-means clustering of meteorological variables. Wind ramp events were then identified based on predicted power series. Experiments showed that dynamic weather made wind power and ramp prediction difficult and thus forecasts had lower accuracy whereas stable weather allowed forecasts with higher accuracy, evidencing that the proposed strategy can provide the information of weather types to electrical grid operators, together with the expected corresponding forecast accuracy under that weather pattern.