为了更精确地预测SO2排放质量浓度，解决非线性随机预测问题，提出了一种基于随机森林特征选择的GWONBEATS算法。通过随机森林算法筛选输入参数的特征，使用灰狼优化算法对NBEATS算法的超参数进行优化；与长短期记忆网络（Long ShortTerm Memory, LSTM）、门控循环神经网络（Gated Recurrent Unit，GRU）以及NBEATS算法对比分析，验证了GWONBEATS算法的有效性。将本算法应用于某大型电网公司大数据平台，探索了复杂智能算法在大数据平台上开展污染物排放预测的可行性。研究结果表明，相较于长短期记忆网络、门控循环神经网络和NBEATS方法，GWONBEATS算法预测误差更小， 其中平均绝对百分比误差MAPE为1.50%，相对均方误差RMSE为0.42，平均绝对误差MAE为0.33，决定系数R2为0.97。

To predict SO2 emission mass concentration more accurately and to solve the nonlinear stochastic prediction problem, a novel grey wolf optimization (GWO) deep learning architecture NBEATS algorithm based on random forest feature selection was proposed. The features of the input parameters were screened by the random forest algorithm, and the hyperparameters of the NBEATS model were optimized using the GWO algorithm; the effectiveness of the proposed algorithm was verified by comparing it with long shortterm memory network (LSTM), gated recurrent unit (GRU) and NBEATS. The algorithm was applied to a large power grid company′s big data platform to explore the feasibility of complex intelligent algorithms to carry out pollutant emission prediction on a big data platform. The results show that the GWONBEATS algorithm has less error compared to LSTM, GRU and NBEATS methods, where MAPE is 1.50%, RMSE is 0.42, MAE is 0.33, and R2 is 0.97.

TK284

国家自然科学基金资助项目(51506052)