Weighted Evaluation of Wind Power Forecasting Models Using Evolutionary Optimization Algorithms

The unpredictability of renewable sources of energy especially wind power causes large fluctuations in the power output. The fluctuations are smoothened by building large amounts of battery storage and/or power reserve capacity. By improving the forecasting accuracy, these reserves can be reduced. We revisit the problem of short-term wind power prediction using statistical and machine learning based modeling techniques. In prior work, we developed a fusion evaluation index to rank various forecasting models. We used eight forecasting models selected from literature and seven evaluation indexes in that study. Each evaluation index was weighted in two parts - an objective normalized weight based on maximizing deviations and a subjective (expert) weight. In this paper, we use two evolutionary optimization algorithms to optimize the objective weights of the indexes. Particle Swarm Optimization (PSO) and Differential Evolution (DE) are used to produce an optimal weight strategy for the six of the seven indexes using a training data set. The weighted objective indexes are then applied to a test dataset with promising initial results. The simulation is based on seven months of actual data from a wind farm in Shanxi province, with a sampling interval of 5 minutes.