Research On Short-term Power Load Forecasting Based On Improved BI-LSTM And CEEMDAN Combined Model

Short-term Power Load forecasting is a crucial basis for grid operation and business management,and its accuracy is important to ensure safe grid operation,reduce production costs and improve economic efficiency.Under the background of the 14 th Five-Year Plan,the demand for future electricity will be increasing,which makes the short-term load forecasting for future electricity consumption more accurate.Therefore,this thesis investigates a combined short-term load forecasting model of deep learning based improved bidirectional long and short term memory neural network(DBI-LSTM)combined with attention mechanism(AT),improved whale optimization algorithm(IWOA)and alignment Permutation entropy(PE)improved complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN),and the main research work is as follows:(1)Through the statistics of electricity load data information of a certain region in China,we analyzed the cyclical characteristic changes of electricity load sequence and analyzed the influence of various factors such as meteorology,season,and date type that exist on electricity load.(2)Through a longitudinal study of deep learning neural networks,a model of deep bidirectional long and short term memory neural network combined with attention mechanism(DBI-LSTM-AT)is constructed.And the DBI-LSTM-AT model is trained using a regional load dataset in China,and validated with different neural network models for comparative analysis(3)The DBI-LSTM-AT short-term Load forecasting model is optimized by IWOA,and the WOA global search capability is first enhanced by using improved nonlinear convergence factor,adaptive probability,adaptive weight factor and random difference variance strategy in WOA.The IWOA-DBI-LSTM model is obtained after the validity of IWOA is verified by using the ICEC general test function,and is compared and analyzed with the model optimized by different algorithms to verify the prediction ability of the constructed model.(4)The CEEMDAN decomposition algorithm is used to decompose the charge data for the characteristics of nonlinearity and non-smoothness of the charge data.The stop iteration criterion of PE improved CEEMDAN is applied to associate IWOA-DBiLSTM-AT model to construct the combined short-term load prediction model.Using the same input data,a validation comparison analysis is performed with different combined models.Finally,to verify the generalization ability and superiority of the model,the model is trained by using a regional electric load dataset in Australia,and the comparative analysis is validated by using different models for a day load and a week load on different seasonal test sets.