Research On Short-term Industrial Load Forecasting Based On Deep Learning

Short-term forecasting of power load is an important means for power system dispatch and detection of power system status.Quick and accurate short-term forecasting of power load is a powerful guarantee for reliable,safe,economical and stable operation of the power system.In2020,industrial electricity will account for 67.0% of the country’s total electricity consumption.As a major industrial electricity consumer in the region,industrial parks have a significant impact on the load curve of the local power grid.As the largest category of electricity consumption,70% of industrial loads are concentrated in industrial parks,which has great potential to participate in the demand side response of the power grid.However,industrial load data has the characteristics of strong volatility,many missing data,and opaque data.Combining the characteristics of the park’s load,selecting suitable methods for accurate short-term load forecasting in the industrial park will not only benefit the adjustment of the regional power grid operation mode,but also provide an important basis for the economic and environmental protection of the grid.However,the current academic forecast of industrial load is still in its infancy.Based on this,the paper selects the short-term forecast of industrial load as the subject for research.The thesis first analyzes the characteristics of industrial load.Aiming at the characteristics of poor data quality and high noise,the paper establishes an EMD noise suppression-AR repair data preprocessing model to effectively improve the data quality and provide guarantee for the subsequent prediction links;for industry With the characteristics of non-linear load and strong volatility,the particle swarm optimization artificial neural network model integrated into the harmony search algorithm is selected for short-term prediction.Artificial neural network has excellent ability to construct nonlinear and complex relationships,but the randomness of its initial weight and set threshold will lead to poor convergence of training results and easy to fall into local optimum.Therefore,the paper uses the global search capability of the particle swarm algorithm to optimize the weights and thresholds of the neural network.At the same time,the particle swarm is used as the harmony memory bank of the harmony search algorithm.The organic combination of the harmony search algorithm and the particle swarm algorithm can be Significantly improve the optimization ability of the particle swarm algorithm,further improve the learning speed of the network,and optimize the network performance.The calculation example analysis part of Chapter 5 takes smelting-based industrial parks as the research object.After analyzing its typical load curve characteristics,the PSOIHS-ANN model established in the paper is used to realize the data pre-processed by EMD denoising and AR repair.The short-term load forecasting is compared with the algorithms that have not passed the EMD noise suppression model,the traditional ANN algorithm,and other popular algorithms.The prediction performance of the EMD-PSOIHS-ANN model established in the paper is demonstrated from multiple angles.The actual measurement results show that the EMDPSOIHS-ANN model established in this paper guarantees the calculation speed,and both MAPE and RMSE are the best;compared to the best-performing LSTM algorithm among other intelligent algorithms,the prediction model MAPE established in the paper An increase of6.541%.The research results of the thesis have certain practical value and academic significance for improving the accuracy and reliability of the short-term forecast of industrial load.