短期电力负荷预测是电力市场经济、电力系统安全运行及调度的重要手段，精准的短期电力负荷预测不仅可以协助相关部门与企业合理安排机组开停、制定购电计划等，还可以向用户提供有价值的用电建议。短期电力负荷的变化具有整体周期和局部随机的“双重特性”，并受到气象等多种变量的影响，单一预测方法难以捕捉其变化规律。因此，本文提出结合负荷分解预测与多变量神经网络预测的组合优化方法，将不同模型从不同角度所获取到的负荷预测信息进行综合，提高短期电力负荷的预测精度与稳定性。

       从负荷自身变化规律的角度出发，采用负荷序列分解的处理方式进行短期电力负荷预测。首先利用改进的奇异谱分析分解方法（ISSA）将电力负荷分解为趋势负荷和随机负荷，然后针对趋势负荷的平稳性和周期性选择季节性差分自回归移动平均模型（SARIMA）进行预测，针对随机负荷所具有的非线性特征采用长短期记忆神经网络模型（LSTM）进行预测，最后将两类模型预测值叠加构成最终负荷预测值，由此构建基于ISSA-SARIMA-LSTM（ISSL）的短期电力负荷预测方法，该方法可以对短期电力负荷变化进行精确的刻画，提高负荷预测稳定性。

       从影响负荷变化的外界因素出发，结合与负荷变化密切关联的多个环境变量进行短期电力负荷预测。首先分析各环境变量对负荷的影响，然后通过分辨系数可变的灰色关联度方法分析各变量与负荷变化的相关关系，构建多变量输入的LSTM神经网络预测模型（M_LSTM），通过模型训练映射各类环境变量与负荷变化之间的隐藏的、非线性关系，提高短期电力负荷的预测精度。

       为降低短期电力负荷的预测风险和提高短期电力负荷的预测精度，建立了基于诱导有序加权平均算子（IOWA）的变权重组合优化预测模型。变权重组合针对ISSL和M_LSTM模型在不同角度上所获得的预测信息差异，按照两种方法在不同点的预测精度进行有序赋权，从而实现不同方法的优缺点互补。实验数据表明，相比于倒数方差法等不变权重组合模型，变权重组合优化能更全面地反应短期电力负荷的变化规律，能有效提高预测模型的稳定性及实际应用能力。

    Short-term load forecasting(STLF) is an important means of power market economy, safe operation and dispatching of power system. Accurate STLF can not only help relevant departments and enterprises to reasonably arrange the start and stop of units and formulate power purchase plans, but also provide valuable suggestions for users. The change of short-term load has the dual characteristics of global cycle and local random, and is affected by many variables such as meteorology. It is difficult to capture the change rule by a single prediction method. Therefore, this paper proposes a combined optimization method combined with load decomposition prediction and multivariable neural network prediction, which integrates load forecasting information obtained from different models from different angles to improve the accuracy and stability of STLF.

    From the perspective of load variation, STLF is carried out by load sequence decomposition. Firstly, the improved singular spectrum analysis decomposition method(ISSA) is used to decompose the power load into trend load and random load. Then, the seasonal difference autoregressive moving average model(SARIMA) is selected to predict the stability and periodicity of the trend load, and the long short-term memory neural network model(LSTM) is used to predict the nonlinear characteristics of the random load. Finally, the predicted values of the two models are superimposed to form the final load forecasting value, and the STLF method based on ISSA-SARIMA-LSTM(ISSL) is constructed. This method can accurately characterize the short-term load changes and improve the stability of STLF. 

    Starting from a variety of external factors affecting load change, STLF is carried out by combining multiple environmental variables closely related to load change. Firstly, the influence of various environmental variables on the load is analyzed, and then the correlation between variables and load changes is analyzed by the grey correlation method with variable resolution coefficients. The multi-variable input LSTM neural network prediction model(M_LSTM) is constructed. The hidden and nonlinear relationship between various environmental variables and load changes is mapped through model training to improve the prediction accuracy of STLF.

    In order to reduce the prediction risk of STLF and improve the prediction accuracy of STLF, a variable weight combination optimization prediction model based on induced ordered weighted averaging operator(IOWA) is established. According to the difference of prediction information obtained by ISSL and M_LSTM models from different angles, the variable weight combination weights model orderly according to the prediction accuracy of the two methods at different points, so as to realize the complementary advantages and disadvantages of different methods. The experimental data show that compared with the constant weight combination model such as the reciprocal variance method, the variable weight combination optimization can more comprehensively reflect the variation law of short-term load, and effectively improve the stability and practical application ability of the prediction model.