- 學位論文
經驗模態分解和改進鯨魚演算法最佳化LSTM模型：以電力負載預測為例
Optimizing LSTM Model with Empirical Mode Decomposition and Improved Whale Optimization Algorithm: A Case Study in Electric Load Forecasting
摘要
隨著智能電網的發展，準確的電力負載預測已變得越來越關鍵，因為它協 助電力公司最佳化負載調度並減少過度發電。本研究選用類神經模型中的長短 期記憶（Long Short-Term Memory, LSTM）模型進行電力負載預測。使用經驗模 態分解（Empirical Mode Decomposition, EMD）作為時間序列資料的預處理技術， 而改進的鯨魚演算法（Improved Whale Optimization Algorithm, IWOA）則是超參 數最佳化的方法。先前對鯨魚演算法(Whale Optimization Algorithm, WOA)的研 究指出，儘管其局部搜索性能良好，但其全域探索能力不足。本研究融合其他 啟發式演算法的特點，以解決 WOA 的限制，並引入其他修改以增強其局部搜 索能力。對 GWO、WOA、PSO、GA、DE 和 IWOA 在 23 個基準函數及 CEC C06 2019 的最佳化性能進行了比較。使用某大樓電力消費數據作為案例研究， 比較網格搜索(Grid Search)、WOA 和 IWOA 在 LSTM 的超參數最佳化表現，使 用 MSE、MAE、R 2作為評估指標。通過應用 EMD 進行資料分解和平滑化，對 比了 WOA 和 IWOA 在 EMD-LSTMs 和 LSTM 的訓練性能。在啟發式演算法的 測試中，IWOA解決了 WOA全域最佳值搜索的問題，並增強了其局部極值探索 能力。IWOA 的區域開發、區域勘探、綜合表現均優於與其他啟發式演算法。 IWOA 可以在較少的執行次數下實現卓越的超參數最佳化效果。在 80 次執行次 數下， IWOA 與 WOA 相比其 MSE 下降 9.21%、MAE 下降 6.76%，而R 2提升 0.16%。EMD 作為預處理方法使得 LSTM 能夠更好地捕捉時間序列資料的複雜 特性。IWOA-EMD相較於 IWOA，MSE下降 70.4%、MAE下降 39.4%、R 2提升 了 0.70%。相較於網格搜索，IWOA-EMD 僅使用 80 次的執行次數就可以超越網 格搜索 400 次執行次數的 MSE、MAE、R 2表現，並且降低其訓練時間 84.0%。 最終在所有執行次數的實驗下，IWOA-EMD 在評估指標上優於其他模型組合。 代表 IWOA-EMD 相較於其他較模型更適合用於負載預測。
並列摘要
With the advancement of smart grids, accurate electric load forecasting has become increasingly critical, as it aids power companies in optimizing load scheduling and reducing overproduction. This study utilizes the Long Short-Term Memory (LSTM) model from neural networks for electric load prediction. The Empirical Mode Decomposition (EMD) is employed as a preprocessing technique for time series data, while the Improved Whale Optimization Algorithm (IWOA) is adopted for hyperparameter optimization. Prior studies on Whale Optimization Algorithm (WOA) have highlighted that, despite its commendable local search capabilities, it falls short in global exploration. This research integrates attributes from other heuristic algorithms to address the limitations of WOA and introduces modifications to bolster its local search prowess. A comparative analysis of the optimization performance of GWO, WOA, PSO, GA, DE, and IWOA across 23 benchmark functions and the CEC-C06 2019 dataset is presented. Using electricity consumption data of a certain building as a case study, we compared the performance of Grid Search, Whale Optimization Algorithm (WOA), and Improved Whale Optimization Algorithm (IWOA) in optimizing hyperparameters for Long Short-Term Memory (LSTM). Mean Squared Error (MSE), Mean Absolute Error (MAE), and R 2 were employed as evaluation metrics. Through the application of Empirical Mode Decomposition (EMD) for data decomposition and smoothing, we contrasted the training performance of WOA and IWOA in EMD-LSTMs and traditional LSTMs. In the testing of heuristic algorithms, IWOA addressed the global optimum search issue of WOA and enhanced its capability for exploring local extreme values. IWOA demonstrated superior regional development, regional exploration, and overall performance compared to other heuristic algorithms. III IWOA achieved excellent hyperparameter optimization with fewer execution cycles. In 80 execution cycles, IWOA exhibited a 9.21% decrease in MSE, a 6.76% decrease in MAE, and a 0.16% increase in R 2 compared to WOA. The use of EMD as a preprocessing method allowed LSTM to better capture the complex characteristics of time series data. IWOA-EMD, compared to IWOA alone, showed a 70.4% decrease in MSE, a 39.4% decrease in MAE, and a 0.70% increase in R 2 . In comparison to Grid Search, IWOA-EMD surpassed the performance of MSE, MAE, and R 2 with only 80 execution cycles, outperforming Grid Search with 400 execution cycles, and reducing training time by 84.0%. Ultimately, in all execution cycles, IWOA-EMD outperformed Grid Search, WOA, IWOA, and WOA-EMD in evaluation metrics, indicating that IWOA-EMD is more suitable for load forecasting compared to other models.