基于最小二乘孪生极限学习机的水电系统发电能力预测方法

Prediction method for hydropower system generation capacity based on least squares twin extreme learning machine

摘要

【目的】针对传统水电发电能力预测精度低、稳定性差等问题。【方法】提出了耦合模态分解、机器学习和群体智能的水电系统发电能力混合预测模型。首先，利用逐次变分模态分解法(SVMD)对原始出力序列进行分解降噪，提取出多尺度特征信号进行分类建模；随后，采用最小二乘孪生极限学习机(LSTELM)对各分解信号进行预测建模，同时运用改进灰狼优化算法(IGWO)对模型参数进行优化，以提升模型的预测性能；最后对各子序列预测结果进行集成，叠加得到最终的预测结果。【结果】结果显示：所提方法在三个水电站中的预测结果精准可靠。在池潭水电站中，预见期为1 d时，所提模型在直接策略和多输入多输出策略中预测结果的纳什系数(NSE)指标较极限学习机模型分别提高了12.88%和12.11%。预见期由1 d增长至8 d时，传统方法预测结果的NSE指标由0.884 0和0.888 5逐渐降低到0.573 5和0.567 1,而本文所提两种策略预测结果分别由0.997 9和0.996 1逐渐降低到0.942 3和0.928 6。【结论】结果表明：所提模型在复杂水电系统发电能力预测中具有较强的稳定性和泛化能力，SVMD有效降低了发电能力序列的噪声影响，最小二乘法和孪生结构提升了LSTELM模型的泛化能力，SVMD-IGWO-LSTELM模型在水文特性稳定的水电站预测精度更高，在水文特性复杂的水电站预测能力略有下降但依旧保持高精度，为变化环境下水电系统发电能力预测提供有效方法。

Abstract

[Objective] To address the issues of low accuracy and poor stability in traditional hydropower generation capacity forecasting, [Methods] a hybrid prediction model for power generation capacity of hydropower systems, combining coupled mode decomposition, machine learning, and swarm intelligence was proposed in this paper. First, the original output sequence was decomposed and denoised using successive variational mode decomposition(SVMD) to extract multi-scale feature signals for classification and modeling. Subsequently, a least squares twin extreme learning machine(LSTELM) was employed to predict each decomposed signal, and an Improved Grey Wolf Optimization(IGWO) algorithm was used to optimize the model parameters, enhancing the prediction performance. Finally, the prediction result of each sub-sequence were integrated and aggregated to obtain the final result. [Results] The result demonstrated that the proposed method achieved accurate and reliable predictions for three hydropower stations. At the Chitan Hydropower Station, for 1-day forecast period, the proposed model improved the Nash-Sutcliffe efficiency(NSE) compared to the extreme learning machine model by 12.88% and 12.11% for direct and multi-input multi-output strategies, respectively. As the forecast period increased from 1 to 8 days, the NSE of the traditional method gradually decreased from 0.884 0 and 0.888 5 to 0.573 5 and 0.567 1, while the NSE of the two proposed strategies decreased from 0.997 9 and 0.996 1 to 0.942 3 and 0.928 6. [Conclusion] The findings indicate that the proposed model is highly stable and generalizable in predicting the generation capacity of complex hydropower systems. SVMD effectively reduces noise influence in the generation capacity sequence, and the least squares method and twin structure enhance the generalization ability of the LSTELM model. The SVMD-IGWO-LSTELM model demonstrates higher prediction accuracy for hydropower stations with stable hydrological characteristics, while prediction accuracy slightly decreases for stations with complex hydrological features, but still remains high. This model provides an effective approach for predicting hydropower system generation capacity under changing environmental conditions.

关键词

逐次变分模态分解法 / 发电出力 / 最小二乘孪生极限学习机 / 改进灰狼优化算法 / 影响因素

Key words

successive variational mode decomposition method / power generation output / least squares twin extreme learning machine / Improved Grey Wolf Optimization algorithm / influencing factors

基金资助

国家电网有限公司重大科技项目“面向重大天气过程的电力系统源荷预测及调节能力评估关键技术研究”(4000-202355381A-2-3-XG)

智能电网国家科技重大专项(2030)

“百兆瓦级全功率变速抽水蓄能机组关键技术及装备”(2024ZD0801600)