Short-term probabilistic predictions of wind multi-parameter based on one-dimensional convolutional neural network with attention mechanism and multivariate copula distribution estimation

Wind speed forecast can effectively guide power grid to schedule adjustable sources to smooth wind uncertainty and ensure system stability. But due to the limited regulating range and velocity of complementary supplies, insufficient capacities can't match wind variations completely always leading wind curtailments and wastes. So wind fluctuation scope and change rate predictions are also highly crucial for dispatching to make more thorough deployments. Therefore, this paper introduces turbulence standard deviation and wind variogram to physically depict these two properties and develops probabilistic short-term combination forecast approach for them and wind speed. This method is based on multi-task one-dimensional convolutional neural network including shared layer to extract information criteria-determined input correlations and task layer to fine-tune output accuracies. And attention mechanism is innovatively added for certain samples to better cater for the wind speed-power curve demand. Results indicate the models stably outperform frequently-used competitors for those more important samples. Then multivariate copula method is employed for the joint distribution estimations of forecasts and actual data to generate conditional fluctuation intervals for each parameter. Superior assessments on test set confirm the validity and generalization of this approach which can provide reliable probabilistic manifold information for adjustable power scheduling.