Very Short-term Interval Prediction Of PV Cluster Power Based On Multi-NWP Feature Extraction

Under the influence of the outstanding economic and environmental benefits of renewable energy,green energy has been applied to electricity production on a large scale.PV has become one of the green energy sources to replace traditional fossil energy generation due to its advantages of easy access and high cleanliness.However,PV power generation is characterized by uncertainty and randomness under the influence of weather,which can pose a threat to the stability of power supply in the power system.Therefore,by analyzing the fluctuation of PV power and the factors affecting the fluctuation to improve the accuracy of PV power prediction can help the power system to better consume PV.Firstly,in order to make the prediction model can be more targeted for learning,the PV power is clustered and divided into the same category for the weather with similar power fluctuation characteristics within a day.Based on the above analysis,the fluctuations of PV power between every two adjacent points were counted,and the power fluctuation anomalies within a day were classified into two types: simple weather and fluctuating weather using the quadratic method.Among them,the power of fluctuating weather is highly random,while the power of simple weather changes gently,and the prediction alone can improve the prediction algorithm’s ability to track the fluctuation process.Secondly,when multiple NWP features are used as inputs to the prediction model along with historical power,too much NWP input to the prediction model may lead to too much redundancy,which in turn degrades the prediction accuracy.In order to include as much information related to PV power as possible without adding burden to the prediction model,this paper uses the parallel factor feature extraction approach to NWP feature extraction for the first time,considering the relevance of NWP features in three dimensions: NWP input variables,very short-term prediction time range and NWP features for feature extraction to ensure the integrity of the input prediction information and avoid the important information loss.The LSTM network with long-term and short-term memory is also used as the PV power very short-term point prediction model for example analysis to further improve the accuracy of prediction.Finally,deterministic point forecasts can only give the predicted power value at a certain moment in time and cannot quantify the impact of forecast deviations on the dispatching plan.In order to further cooperate with the normal operation of power system scheduling,interval prediction of PV clusters can better help power system scheduling to make reasonable decisions.The spatial and temporal characteristics of different PV plants are considered comprehensively,and the K-Modes model,which is more suitable for discrete data clustering,is used to divide plants with similar spatial and temporal characteristics into a sub-cluster,so as to smooth the forecast errors of different PV plants within the sub-cluster.The clustered subclusters are point predicted using LSTM,and the total cluster point prediction results are obtained using the cumulative method.The kernel density estimation method is applied to the interval prediction of the total cluster point prediction results to meet the requirements of safety and economy of the power system after large-scale PV grid connection.