Research On Short-term Power Forecasting Method Of Wind Power Cluster Based On Deep Learning

Wind power prediction is one of the critical technologies for the power system to deal with wind power volatility.In the scenario of large-scale wind power integration,accurate station-level,especially regional-level short-term wind power prediction is of great significance to the safe and economic dispatch of the power system.For station-level forecasting systems,generally single-source and single-location Numerical Weather Prediction(NWP)is used for short-term wind power forecasting.However,single-source NWP data has limited adaptability to complex weather conditions,and single-location NWP data cannot characterize the Spatio-temporal coupling relationship of the natural wind fluctuation process in large wind farms.For regional-level forecasting systems,the conventional "one wind farm,one forecasting" method ignores the meteorological contact and spatio-temporal correlation characteristics between wind farms.In response to these problems,the author studies the station-level and regional-level short-term wind power prediction methods based on deep learning,using the characteristics of multiple wind turbines and wind farms as input,fully excavating the spatio-temporal variation of wind power and improving the prediction accuracy.The main research contents are as follows:1)Research on correction method of the multi-location NWP in wind farm based on multi-to-multi mappingOn the basis of exploring the distribution law of multi-source and multi-location NWP errors in a wind farm,a multi-location NWP correction model in a wind farm based on multi-to-multi mapping is established.The model is built through deep belief network,using multi-source and multi-location NWP data as input and the measured wind speed data at multiple wind turbines as output—realizing simultaneous correction of the multi-location NWP wind speed in the wind farm.It improves the adaptability to weather conditions and can more accurately describe the flow correlation of the flow field in the wind farm.The results of calculation examples show that the NWP wind speed correction effect of the constructed model is better than that of single-source multi-to-multi mapping,multi-source single-unit model and single-source single-unit model.2)Research on short-term wind farm power prediction method based on multi-location NWP feature screeningFirst,for the power prediction of a single wind turbine,the multi-location original and modified NWP data in the wind farm are used as the initial and basic characteristic parameters.The Gradient Boosting Decision Tree(GBDT)is used to carry out feature screening,and then retrain the model for each wind turbine unit with the filtered features as input,and finally the power prediction results of wind turbines are added as the wind farm power prediction result.After verification and analysis of examples,the proposed method can effectively improve the accuracy of short-term wind farm power prediction and has practical engineering value.3)Research on centralized forecasting method of short-term wind farm cluster power based on Convolutional Long and Short-term Memory NetworkThe power generation output correlation between regional wind farm cluster is analyzed.Based on the Convolutional Long and Short-term Memory Network(ConvLSTM),a multi-to-multi mapping of the day-ahead wind farm cluster short-term power centralized prediction model is established.The model takes the multi-source NWP time series data of the wind farm cluster,the historical measured time series data of the wind farm cluster,and the historical power data of the wind farm cluster in the last time step as input.It can realize the simultaneous prediction of all wind farms in an area.The calculation examples show that the prediction accuracy of the centralized prediction model is higher than that of the split-station modelling model.The historical measured time-series data of wind farms can assist in short-term wind power prediction and correct errors.