LAUSR

For the probabilistic forecasting of photovoltaic generation, the design of a feasible online predictive framework is a challenging problem. To deal with this issue, an ensemble dynamic OS-ELM based on quantile estimation has been proposed. Considering inherent intermittent and random variations of the photovoltaic sequence, the weights of the forecasting model are updated according to the loss suffered. At the same time, the accommodative feature threshold can be calculated on the basis of increasing characteristics space, which builds the optimal network architecture for point forecast. A two-dimensional kernel density estimation algorithm based on a fuzzy inference method is exploited. This method breaks through the assumptive limit of error distribution. The changeable parameters help the model to be suitable for power fluctuations. Both the arriving data sample and the increasing feature space are tackled. Numerical experiments are conducted on two practical applications for solar power systems. The results show that the proposed algorithm not only has lower generalization error, but also provides higher confidence coefficient.