Research On Ultra-short-term Hybrid Forecasting Of Wind Power Based On BN Decomposition

Wind power generation is one of the most mature renewable energy power generation technologies,and it is also the priority development direction of energy policies in the world.However,the output of wind power has the characteristics of strong randomness,volatility and intermittentity.The uncertainty of output leads to the phenomenon of abandoning wind power and limiting power generation from time to time.The access of large-scale wind power poses a severe challenge to the safe and stable operation of power grid.Accurate prediction of wind power power can not only weaken the influence of wind power sequence volatility,but also ensure the safety and stability of wind power operation after it is connected to the power system.Based on the historical data,the power fluctuation characteristics and uncertainty of wind farms are deeply studied.The main factors influencing the ultra short-term wind power prediction and the influence of the uncertain output of wind power on the power grid are analyzed.The evaluation indexes of wind power prediction are elaborated,which lays a theoretical foundation for the large-scale ultra short-term wind power prediction.Aiming at the problems of poor orthogonality and model aliasing in ultra short-term term power prediction preprocessing,a wind power pre-processing method based on BN(Beveridge Nelson)decomposition is proposed.Firstly,ADF is used to test the stationarity of the original series,and the non-stationarity of the original power series is eliminated by first-order difference,and then decomposed into trend components and periodic components.The trend component includes the deterministic component that tends to be stable and the random component.The ARIMA model is used to estimate the deterministic component of the power sequence trend component,and extract the deterministic component and random component.At the same time,in order to eliminate the irregular components of the high frequency in the sequence,the BK filtering method is used to extract the periodic components of the power sequence.The analysis of calculation examples shows that the proposed preprocessing method can effectively eliminate the residual components in the decomposition process,and at the same time weaken the volatility of wind power.In order to reduce the maximum error caused by a single prediction algorithm,a hybrid prediction model is proposed.The hybrid models are independent of each other and generate their own prediction values.Aiming at the stable trend of deterministic components,a recurrent neural network prediction model(GRU)of gating unit is established,which can alleviate the disappearance of gradient and has strong fitting ability;Aiming at the problem of large cycle time span of periodic components,a prediction model of least squares vector machine(LSSVM)is established,which can flexibly adapt to its time series attributes;Aiming at the complex characteristics of random component data changes,establish an extreme learning machine(ELM)prediction model with strong generalization performance and fast convergence speed.In order to clarify the error distribution rule of the combined model and avoid the network parameters from falling into the local optimum,the adaptive neuro-fuzzy inference system(ANFIS)optimized by genetic algorithm(GA)is used to correct the prediction errors.The analysis of calculation examples shows that the hybrid forecasting model proposed in this paper can reduce the maximum error caused by a single forecasting model,and at the same time improve the accuracy of the forecasting algorithm.