Research On Ultra-Short-Term Wind Speed And Solar Radiation Corrected Forecasting Based On Machine Learning And Swarm Intelligence

With the increasingly prominent issue of global warming caused by greenhouse gases and the continuous growth in energy demand,renewable energy,as a clean and sustainable form of energy,has garnered widespread attention worldwide.Wind energy and solar energy are two major sources of renewable energy,known for their wide distri-bution,lack of pollution,and zero carbon emissions during electricity generation,they have become the primary means for countries to achieve net-zero emissions.However,the intermittency,randomness,and volatility of these two energy sources pose signifi-cant challenges to energy management and the operation of power systems.Therefore,improving the forecast accuracy of ultra-short-term wind speed and solar radiation has become a key factor in building a new power system that relies primarily on integrat-ing new energy sources,contributing significantly to China’s goal of achieving"Dual Carbon"at an earlier date.To address the issues of insufficient forecast accuracy and limited forecast stability caused by the strong nonlinear characteristics of wind speed and solar radiation,this pa-per proposes ultra-short-term wind speed and solar radiation correction forecast models using numerical weather prediction models,machine learning and swarm intelligence methods,respectively.Firstly,numerical weather prediction models are used for fore-casting wind speed and solar radiation.Subsequently,artificial neural networks(ANN)and swarm intelligence optimization methods are employed to correct the wind speed forecasts for four different time steps,and chaotic systems are introduced to enhance swarm intelligence optimization,addressing the issue of susceptibility to local optima and further improving forecast accuracy.Then,to meet the power forecast of future large-scale photovoltaic power stations,fully convolutional neural networks(FCN)are used to establish correction forecast models for ultra-short-term solar radiation across four different time steps.Finally,a comparative analysis of the forecast results from dif-ferent models is conducted.The main research results and conclusions are as follows:(1)The WRF model forecasts wind speed at turbine heights with significant errors,with an average Mean Absolute Error(MAE)of 3.19 m/s for the four tested months.Using the Autoregressive Integrated Moving Average(ARIMA)model to correct the forecasts,the average MAEs for the four different time steps in the four months are1.64 m/s,1.87 m/s,2.04 m/s and 2.19 m/s,respectively.Seven different ANN models show varying forecast accuracy in different months but outperform the ARIMA model.The optimal ANN model yield average MAEs of 1.29 m/s,1.66 m/s,1.89 m/s and2.07 m/s for the four different time steps,compared to the wind speed forecasted by the WRF,the percentage improvement of MAEs are 59.56%,47.96%,40.75%and 35.11%,respectively.(2)Compared to the ANN model without data decomposition,the ANN model with data decomposition has improved the accuracy of wind speed forecast to a cer-tain extent.After the Variational Mode Decomposition(VMD),the average MAEs of wind speed corrected by all ANN models for the four different time steps are 0.42 m/s,0.48 m/s,0.57 m/s and 0.68 m/s,respectively,compared to the wind speed forecasted by the WRF,the percentage improvement of MAEs are 86.83%,84.95%,82.13%and78.68%,respectively.After the Wavelet Packet Decomposition(WPD),the correspond-ing average MAEs are 0.43 m/s,0.69 m/s,0.98 m/s and 1.21 m/s,respectively,com-pared to the wind speed forecasted by the WRF,the percentage improvement of MAEs are 86.52%,78.37%,69.28%and 62.07%,respectively.After the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(ICEEMDAN)decom-position,the corresponding average MAEs are 0.59 m/s,0.88 m/s,0.95 m/s and 1.08m/s,respectively,compared to the wind speed forecasted by the WRF,the percentage improvement of MAEs are 81.5%,72.41%,70.22%and 66.14%,respectively.Overall,the wind speed correction results shows the most significant improvement after VMD decomposition,while WPD and ICEEMDAN decompositions yield relatively limited correction results.(3)The combination correction forecast models based on swarm intelligence opti-mization outperform data decomposition-ANN models in terms of forecast accuracy,and the Linear Regression Network(LRN)optimizer performs less effectively than swarm intelligence optimizers,resulting in the poorest forecast accuracy among all optimizers,in contrast,the Chaotic Sparrow Search Algorithm(CSSA)optimizer,as proposed in this paper,achieves the best correction forecast results.LRN combination correction forecast models yield average MAEs of 0.31 m/s,0.47 m/s,0.58 m/s and0.69 m/s for the four different time steps,respectively,compared to the wind speed forecasted by the WRF,the percentage improvement of MAEs are 90.28%,85.27%,81.81%and 78.37%,respectively.CSSA combination correction forecast models yield average MAEs of 0.17 m/s,0.25 m/s,0.31 m/s and 0.38 m/s for the four different time steps,respectively,compared to the wind speed forecasted by the WRF,the percentage improvement of MAEs are 94.67%,92.16%,90.28%and 88.09%,respectively.(4)The WRF-Solar model overestimates solar radiation overall,with an average bias of approximately 33.77 W/m~2and an average MAE of 53.94 W/m~2for the four months.Among the five ANN correction forecast models,FCN performs the poorest,while the other networks yield varying correction forecast results for different months and time steps.The average MAEs of all ANN models for forecasting solar radiation for the four different time steps are 20.96 W/m~2,24.76 W/m~2,28.8 W/m~2and 29.24W/m~2,respectively,compared to the solar radiation forecasted by WRF-Solar,the per-centage improvement of MAEs are 61.14%,54.1%,46.59%and 45.79%,respectively.Among the three optimizer combination correction forecast models,the Chaotic Aquila Optimization(CAO)model proposed in this paper yields the best correction results,the average MAEs for the four different time steps are 14.71 W/m~2,19.19 W/m~2,22.35W/m~2and 24.27 W/m~2,respectively,compared to the solar radiation forecasted by WRF-Solar,the percentage improvement of MAEs are 72.73%,64.42%,58.56%and55.01%,respectively,and the spatial distribution of solar radiation bias is more uniform in the CAO correction forecasts.(5)Whether for single-site wind speed correction forecasts or regional-scale so-lar radiation correction forecasts,convolutional networks demonstrate powerful fea-ture extraction and correction forecast capabilities,furthermore,combining chaotic sys-tems with swarm intelligence optimization methods makes it easier for the population to escape local optima,thus finding optimal combination weights.Experimental re-sults prove that optimizers of this type proposed in this paper exhibit strong robustness and universality,showing good forecast accuracy and stability in ultra-short-term wind speed and solar radiation correction forecasts,with promising practical applications.