Research On Short-term Power Load Forecasting Model Based On ISSA-AM-BiLSTM

The rapid development of contemporary society has led to increasing complexity of power systems.To ensure the economic,safe,efficient and reliable operation of power systems,research in short-term load forecasting is essential.However,due to the nonlinearity of the load affected by multiple factors,it is difficult to achieve accurate short-term load forecasting.Neural networks have excellent nonlinear fitting capabilities,making them a promising approach for shortterm load forecasting.Therefore,research on short-term load forecasting based on neural networks has been carried out,with the following main objectives:To address the complexity,diversity,and regularity of short-term power load data,a short-term load forecasting model was constructed based on BiLSTM.BiLSTM is an extension of LSTM that can remember longer-term information and consider historical information.It can also enable the network to learn future information in advance,which is advantageous for prediction problems.Considering that the large amount of information may lead to long processing time,AM was incorporated into the model to improve its performance by simulating the working principle of the human brain when processing massive information.AM can discover the hidden correlation between data,obtain the contribution ratio of feature information,and weight the data accordingly to highlight the features with greater influence,thus updating the input information.After the model training,the data was propagated downwards to obtain the final model prediction results.In order to overcome the difficulty of selecting hyperparameters for the AMBiLSTM model,intelligent optimization algorithms are used to perform parameter optimization.The SSA algorithm is a relatively excellent new intelligent optimization algorithm,but it still has the problem of being trapped in local solutions and unable to jump out.Additionally,when the population is unevenly distributed,it can affect the convergence speed of the model.Therefore,further improvements are needed for SSA.Firstly,the population initialization method of SSA is replaced with Sin chaos to lay the foundation for global search optimization of the model.Secondly,to make the model’s global search optimization more thorough,the previous generation’s global optimal solution is introduced into the SSA discoverer’s position update strategy.Additionally,dynamic adaptive weights are integrated to speed up the convergence of the model and coordinate the model’s ability to search for local and global optimization.Then,the position update method of the warning role is improved to further enhance the model’s optimization ability.Finally,Cauchy mutation and reverse learning strategies are introduced to SSA,which improves the model’s ability to jump out of local spaces.Additionally,a greedy rule is added to provide a decision basis for the updated disturbance mutation.ISSA is used to optimize the AM-BiLSTM model,and the best fitness value obtained is assigned to the corresponding hyperparameters of the prediction model,which overcomes the problem of difficult hyperparameter selection and improves the accuracy of the model’s predictions.Thirteen prediction models,including BP,LSTM,BiLSTM,PSO-BP,PSOLSTM,SSA-LSTM,AM-BiLSTM,WOA-AM-BiLSTM,PSO-AM-BiLSTM,IPSOAM-BiLSTM,SSA-AM-BiLSTM,ISSA-AM-BiLSTM,and IPSO-LSSVM,were built and compared in single-day,double-day,and seven-day load forecasting experiments with practical examples.The results confirmed the superiority of the ISSA-AM-BiLSTM model.