精确的短期负荷预测能通过预知所需用电负荷来指导电网的调度运行。但电力负荷不仅与用户用电习惯相关,还容易受到温度、湿度等气象因素的影响。因此,文章在现有的使用负荷历史数据基础上,增加了影响区域型电力负荷的气象数据,并考虑高维气象参量数据对预测算法的过拟合问题,提出了基于稀疏核主成分分析(sparse kernel principal component analysis, SKPCA)的高维气象数据降维方法。进而以历史负荷功率和SKPCA重构后主成分为输入,构建了基于卷积神经网络(convolutional neural network, CNN)和长短时记忆(long-short-term memory, LSTM)神经网络的混合深度学习预测模型。CNN-LSTM模型可以同时提取负荷功率及气象数据的空间和时间相关特征,从而全面利用数据的时-空相关性特征,提升负荷功率的短期预测精度。相比于常见的数据降维和负荷预测方法,文章所提方法的数据维度降低了71.43%,预测精度达到了98.92%。结果表明,所提算法融合通过SKPCA降维后的气象数据能够显著提升区域型电力短期负荷预测准确度。

Accurate short-term load forecasting can provide guidance for the dispatching operation of power grids by predicting the required power loads. However, the power load is not only related to the user’s electricity consumption habits but is also easily affected by meteorological factors such as temperature and humidity, Therefore, based on existing historical load data, this paper incorporates meteorological data that affect regional power loads, considers the overfitting problem of high-dimensional meteorological parameter data to the prediction algorithm, and proposes a dimensionality reduction method for high-dimensional meteorological data based on sparse kernel principal component analysis (SKPCA). Subsequently, taking the historical load power and principal components reconstructed by SKPCA as the input, we construct a hybrid deep learning prediction model based on a convolutional neural network (CNN) and a long short-term memory (LSTM) neural network. The CNN-LSTM model can extract the spatial and temporal correlation characteristics of the load power and meteorological data simultaneously to fully utilize the temporal-spatial correlation characteristics of the data and improve the short-term prediction accuracy of the load power. Compared with common methods of data dimension reduction and load forecasting, the data dimensions of this method decrease by 71.43%, and the prediction accuracy reaches 98.92%. The results show that the proposed algorithm can significantly improve the accuracy of regional power short-term load forecasting by fusing meteorological data after dimensionality reduction using SKPCA.

电力负荷预测实质是时间序列预测问题，存在非平稳性和影响因素的复杂性。为了提高预测精度，解决长短期记忆神经网络(LSTM)参数选取随机性大、选取困难的问题，本文提出了一种利用麻雀搜索算法(SSA)优化长短期记忆神经网络参数的短期电力负荷预测模型（SSA-LSTM），通过历史用电负荷数据、相关影响因素数据对待预测日进行负荷预测。首先，对历史用电负荷数据、天气、节假日等影响因素进行预处理。其次，将处理好的数据用以训练模型，借助麻雀搜索算法对长短期记忆神经网络的参数进行寻优，使输入数据与网络结构更好地进行匹配。最后，进行负荷预测同时对比其他算法模型进行分析。算例结果表明，本文所提模型能够有效提高预测精度且在进行短期负荷预测中具有有效性。

Power load forecasting is essentially a time series forecasting problem, which has non-stationarity and complexity of influencing factors. In order to improve the prediction accuracy and solve the problem of large random and difficult selection of long-short term memory (LSTM) neural network parameters, a short-term power load forecasting model (SSA-LSTM) that uses the sparrow search algorithm (SSA) to optimize the parameters of the long-short term memory neural network is proposed. The historical power load data related influencing factor data is used to make load forecasts on the day to be forecasted. First, the historical power load data, weather, holidays and other influencing factors data are preprocessed. Secondly, the processed data is used to train the model, and the parameters of the long-short term memory neural network are optimized with the help of the sparrow search algorithm to better match the input data with the network structure. Finally, load forecasting is performed and other algorithm models are compared for analysis. The results of calculation examples show that the model proposed in this paper can effectively improve the prediction accuracy and is effective in short-term load forecasting.

针对综合能源负荷易受气象因素影响及其异质能量耦合特性所导致的预测建模复杂、准确性不高等问题，提出了一种考虑温湿指数与耦合特征的负荷短期预测模型。首先，在深入挖掘多元负荷耦合特征的基础上，结合温湿指数构造计及多因素影响的输入变量；然后，利用核主成分分析（KPCA）法在确保信息有效的前提下完成对预测输入空间的降维处理，并基于门控循环单元（GRU）神经网络进行预测建模，进一步引入Attention机制实现重要特征的差异化提取；最后，选取某实际系统电、冷负荷数据进行仿真。仿真结果表明，基于KPCA-GRU-Attention模型的电、冷负荷短期预测结果的均方根误差和平均绝对百分误差分别为1 025 kW，2.7%和2 167 kW，2.9%，准确性得到了显著提升。所提方法能够在考虑多因素影响的基础上有效提高综合能源负荷的短期预测精度，实现了对用能需求的精准感知。

To address the vulnerability of integrated energy load prediction to meteorological factors and the complexity and low accuracy of prediction models caused the coupling characteristics of heterogeneous energy， a short-term load forecasting model considering temperature-humidity index and coupling characteristics is proposed. Excavating the coupling characteristics of multiple loads， the input variables considering the influence of temperature-humidity index and multiple factors are constructed. Ensuring that the information is valid，kernel principal component analysis （KPCA） is used to complete the dimensional reduction of prediction input space， and the prediction model is built based on gated recurrent unit （GRU） neural network. Attention mechanism is introduced into the model to extract important differentiated features. Finally， the electric and cooling load data of a practical system are selected for simulation， and the results show that the root mean square errors and mean absolute percentage errors of the electric and cooling load predicted by KPCA-GRU-Attention model are 1 025 kW， 2.7% and 2 167 kW， 2.9 %， respectively. The accuracy has been significantly improved. The proposed model effectively improves the short-term prediction accuracy of integrated energy loads by considering the influence of multiple factors， realizing the accurate perception on energy demand.

精准的多元负荷短期预测是综合能源系统调度和运行的基础。综合能源系统中的多种负荷之间存在较强的耦合作用,目前已有的单一负荷预测难以挖掘不同负荷之间复杂的内在联系。对此,提出一种基于多头概率稀疏自注意力模型的多元负荷短期预测方法。首先,采用皮尔逊相关系数分析多元负荷之间的相关性,并提取多元负荷之间的耦合特征;然后,使用改进位置编码的多头概率稀疏自注意力机制学习长序列输入的依赖关系,并且采用多元预测任务的参数软共享机制,通过不同子任务对共享特征的差异化选择,实现多元负荷的联合预测;最后,在亚利桑那州立大学Tempe校区的多元负荷数据集上对所提模型的性能进行验证,结果表明所提预测方法相较于其他预测模型能够有效提高预测精度。

Accurate short-term forecasting of multienergy loads is the basis for the dispatch and operation of integrated energy systems. There is a strong coupling between multiple loads in an integrated energy system, and the existing single load forecasting is challenging to explore the complex internal relationship between multiple loads. Therefore, a short-term forecasting method for multienergy loads based on a multihead probabilistic sparse self-attention (MPSS) model was proposed. First, the Pearson correlation coefficient was used to analyze the correlation between multiple loads, the coupling features between multiple loads were extracted, a multihead probabilistic sparse self-attention mechanism with improved location coding was used to learn the dependencies of long-sequence inputs, and the parameter soft sharing mechanism of multivariate prediction tasks was adopted. The sharing mechanism realizes the joint prediction of multiple loads through a differentiated selection of shared features using different subtasks. Finally, the performance of the proposed model was verified using the multiple-load dataset of the Tempe Campus of Arizona State University. Compared with other forecasting models, the results show that the proposed multivariate load forecasting method can effectively improve forecasting accuracy.

为了提高短期电力负荷预测的精度,提出一种基于自适应柯西变异粒子群(ACMPSO)算法优化长短期记忆(LSTM)神经网络的短期电力负荷预测模型(ACMPSO-LSTM)。针对LSTM模型参数较难选取的问题,采用ACMPSO算法进行LSTM模型参数寻优,利用非线性变化惯性权重来提高PSO算法的全局寻优能力和收敛速度,并在寻优过程中添加了基于遗传算法中的变异操作,减小粒子陷入局部最优解的风险。仿真结果表明,ACMPSO优化LSTM的方法能够有效提高短期电力负荷预测的精度和稳定性。

To improve the accuracy of short-term power load forecasting, a short-term power load forecasting model (ACMPSO-LSTM) based on long-short memory neural network (LSTM) optimized by adaptive Cauchy mutation particle swarm optimization (ACMPSO) is proposed. <em>For the problem of difficult selection of LSTM model parameters, ACMPSO is used to optimize model parameters, and non-linear changing inertia weights are adopted to improve the global optimization ability and convergence speed of PSO algorithm. In the optimization process, a mutation operation based on genetic algorithm is added to reduce the risk of particles falling into local optimal solutions. </em>The simulation results show that the ACMPSO algorithm for LSTM can effectively improve the accuracy and stability of short-term power load forecasting.