米浩然, 申森林, 卫紫任,等.改进麻雀搜索算法优化的短期电力负荷CNN-BiLSTM融合预测模型[J].河南理工大学学报（自然科学版）,2026,45(5):107-115.

MI H R, SHEN S L, WEI Z R,et al.An improved sparrow search algorithm （SSA）-optimized CNN-BiLSTM hybrid model for short-term electric load forecasting[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(5):107-115.

改进麻雀搜索算法优化的短期电力负荷CNN-BiLSTM融合预测模型

米浩然, 申森林, 卫紫任, 李巍, 李亚檑

国网河南省电力公司 焦作供电公司，河南 焦作 454000

摘要: 目的为提高电力负荷预测精度，降低预测难度，进行改进麻雀搜索算法优化的短期电力负荷CNN-BiLSTM融合预测模型研究。  方法 以河南省某地区110 kV变电站2022年的电力负荷数据为基础，提出一种基于改进麻雀搜索算法优化的卷积神经网络-双向长短期记忆网络（ISSA-CNN-BiLSTM）融合预测模型。首先，以日期信息、气象数据和历史负荷数据等构建特征集作为模型的输入变量，并对数据的异常值、缺失值和数据不规范等问题进行预处理；其次，采用CNN模型有效提取特征图中连续/非连续数据之间的潜在联系并构造时序序列的特征向量，再基于BiLSTM模型对其进行训练及预测；同时利用正余弦策略和柯西变异策略对传统的SSA进行强化，实现对模型超参数的自动寻优，得到预测结果。  结果仿真结果表明，与LSTM、PSO-LSTM、CNN-LSTM和Transformer-BiLSTM预测模型相比，ISSA-CNN-BiLSTM模型的预测误差MAPE降低为1.570 3%，平均绝对误差为0.427 6，均方根误差为0.645 3，预测精度得到明显提高。  结论 ISSA-CNN-BiLSTM融合预测模型能够克服负荷预测随机因素的影响，具有更强的负荷时间相关性捕捉能力和泛化能力，可以为实际电力系统的高效运行和优化调度提供新的思路和方法。

关键词:短期负荷预测;CNN-BiLSTM融合模型;改进麻雀搜索算法;预测精度

doi:10.16186/j.cnki.1673-9787.2025120085

基金项目:国网河南省电力公司科技项目（5217C0250001）

收稿日期:2025/12/26

修回日期:2026/04/28

出版日期:2026-07-24

An improved sparrow search algorithm （SSA）-optimized CNN-BiLSTM hybrid model for short-term electric load forecasting

Mi Haoran, Shen Senlin, Wei Ziren, Li Wei, Li Yalei

Jiaozuo Power Supply Company， State Grid Henan Electric Power Company， Jiaozuo 454000， Henan， China

Abstract: Objectives Accurate power load forecasting plays a crucial role in ensuring stable power system operation and rational resource allocation. However， with the increasing penetration of renewable energy and the influence of complex meteorological conditions and diversified load structures， the uncertainty and difficulty of load forecasting have further increased. Methods Based on the electric load data of a 110 kV substation in Henan Province in 2022， an ISSA-CNN-BiLSTM hybrid forecasting model optimized by an improved sparrow search algorithm （ISSA） was proposed. First， a feature set was constructed using date information， meteorological data， and similar-day load data， and data preprocessing was performed to handle outliers， missing values， and inconsistent formats. Then， a CNN was employed to extract spatial–temporal features and construct feature vectors， which were subsequently input into a BiLSTM network for training and prediction. In addition， sine-cosine and Cauchy mutation strategies were introduced to enhance the traditional SSA for automatic hyperparameter optimization.  Results Compared with LSTM， PSO-LSTM， CNN-LSTM， and Transformer-BiLSTM models， the proposed ISSA-CNN-BiLSTM model achieved the lowest forecasting error， with a mean absolute percentage error （MAPE） of 1.570 3%， mean absolute error （MAE） of 0.427 6， and root mean square error （RMSE） of 0.645 3， demonstrating significant improvement in forecasting accuracy.  Conclusions The proposed ISSA-CNN-BiLSTM hybrid model effectively reduces the impact of stochastic factors in load forecasting. It enhances the ability to capture temporal dependencies and improves generalization performance， providing a new approach for efficient power system operation and optimal scheduling.

Key words:short-term load forecasting;CNN-BiLSTM hybrid model;improved sparrow search algorithm;prediction accuracy