Fig 1.
TCN Model.
Fig 2.
DTR data forecasting process.
Table 1.
Line parameters.
Fig 3.
Improved VMD decomposition of DTR data.
Legend: IMF1-IMF4 = intrinsic mode functions from improved VMD; Res = residual component; X-axis = time (h); Y-axis = DTR value (A).
Fig 4.
Optimization performance comparison results.
Fig 5.
Optimization performance of different algorithms.
Legend: Algorithms compared: SMA (Smile Mould Algorithm). PSO (Particle Swarm Optimization), GA (Genetic Algorithm); Y-axis = Fitness Function Value; X-axis = Iteration Count.
Fig 6.
Comparison of prediction results between the proposed model and deep time series modeling. model.
Legend: “proposed model” = improved VMD-time-varying multi-model ensemble; deep time-series: LSTM, GRU, Transformer; X-axis = Sample point; Y-axis = predicted DTR (A).
Fig 7.
Comparison of prediction results between the proposed model and traditional learning models.
Legend: “proposed model” = improved VMD-time-varying multi-model ensemble; Other models: BP (Back Propagation), LSSWM (Least· Squares Support Vector Machine); X-axis = Sample point; Y-axis = predicted DTR (A).
Fig 8.
Comparison of evaluation metrics results between the proposed model and other models.
Legend: Metrics: R2 (determination coefficient, %), RMSE (Root- Mean Square Error), MAE (Mean Absolute Error), MAPE (Mean Absolute Percentage Error, %); Models: improved VMD-time-varying multi-model ensemble, TCN, LSSVM, BP, LSTM, Transformer, GRU.
Fig 9.
DTR prediction results.
Fig 10.
Evaluation metrics results.
Table 2.
Comparison results.
Table 3.
Sensitivity analysis results of SMA algorithm parameters.
Table 4.
Sensitivity analysis results of VMD hyperparameters.
Table 5.
Sensitivity analysis results of VMD hyperparameters.
Table 6.
Sensitivity analysis results of dynamic weighted sliding window length.