Conv-TasNet運用於家庭電力訊號分解

Abstract

台灣為海島國家，能源仰賴進口，隨著用電需求增加，節約電力成為重要的課題。以往著重於工商業等高壓用戶，透過建立完善的監控系統，進行各種能源的掌握，協助電力改善及設備能效管控。台灣的低壓用戶佔全體用戶89%，因此電能議題上不可忽視。對低壓用戶而言，使用大量的監控設備進行能源管控，不僅增加成本同時負擔監控設備能耗，反而造成浪費。本研究為解決電力系統監控及設備管控問題，研究將單一電力訊號，分離成多種設備個別訊號之模型。 單一電力訊號分離已發展多年，隨著運算效能提升深度學習被廣泛運用於該領域。有鑑於語音分離領域，在時域領域能夠有所突破。本研究將語音分離模型 Conv-TasNet 轉換領域，透過 REDD 資料集測試模型轉換之參數及架構調整。研究 Conv-TasNet 應用於電力分離任務之可行性以及成效。 實驗結果顯示，以往使用SI-SNR作為目標函數，電力訊號領域時改用MSE函數以及在輸出層加上ReLU激活函數能夠提高模型成效。該模型於五種設備分離時，分離訊號於總電力訊號佔比，相對誤差小於10%。透過家庭一到三訓練模型，測試於家庭五之電冰箱及微波爐，分離訊號之MAE誤差約為24及5.6瓦特。

Taiwan is an island country, energy using in Taiwan depends on imports. Through the increase in electricity demand, power saving has become an important task. Previously, it focused on high-voltage users such as industry and commerce. To build a complete monitoring system, it can control various energy sources to assist in power improvement and equipment energy efficiency control. But Taiwan’s low-voltage users account for 89% of all users, therefore the electrical connection cannot be ignored. For low voltage customers, the use of a large number of monitoring equipment for energy management and control causes waste while increase the cost and bear the consumption of monitoring equipment at the same time. In order to solve the problems of power system monitoring and equipment management control, this research build a model to separate the single power signal into multiple devices. Single power signal separation has been developed for many years, and deep learning has been widely used in this field with the improvement of computing performance. In view of the field of speech separation, a breakthrough can be made in the time domain. In this research using the speech separation model Conv-TasNet transform to the field. Through the REDD data set, testing model conversion parameters and structure adjustments used to study the feasibility and effectiveness of Conv-TasNet, applied to power separation tasks. According to the results, SI-SNR was used as the objective function in the past. The MSE function was used in the power signal field and the ReLU activation function was added to the output layer were the most effective of the model. Using this model to separate the five devices, the separation signal accounts for the proportion of the total power signal its relative error is less than 10%. Through the family one to three training model, tested in household five refrigerators and microwave ovens, the MAE error of the separated signal is about 24 and 5.6 watts.