透過鯨魚演算法調變權重之類神經網路應用於三電力電動車輛系統上最佳化能源管理策略

Abstract

本研究旨在透過類神經網路應用於三電力電動車輛能量管理策略上，以增進未來可實現於車輛控制器硬體上之可行性。類神經網路之訓練數據集，由混合動力車輛之能量管理領域上，常應用之最小等效能耗法(Equivalent Consumption Minimum Strategy, ECMS)進行收集。首先，本研究先建立具有三電力源電動車輛系統特性之低階動力學模型，其中包含電力源(燃料電池、鋰電池、超級電容)，所需相關參數皆從商業軟體Advanced Vehicle Simulator(ADVISOR)獲得，基於目標車型 Toyota Mirai 設計具有三電力架構之增程式版本，動力系統包含110kW燃料電池組、40Ah鋰三元電池組以及165F/48V超級電容組，並搭配150kW交流馬達。作為訓練數據集來源，ECMS控制策略為一包含六階層環狀結構，分別為鋰電池健康度、需求功率、電池殘電量、超級電容殘電量、燃料電池與需求功率比值以及鋰電池與需求功率比值，作為訓練類神經網路使用，輸入為ECMS前四層參數，輸出則為後兩層：燃料電池與需求功率比值與鋰電池與需求功率比值，本研究透過鯨魚仿生演算法針對類神經網路調變權重(WOA-ANN)與傳統倒傳遞法(BPANN)進行分析比較。為評估類神經網路之效益，本研究同時也建立規則庫控制法。為分析能耗差異性，透過新歐洲駕駛循環(New European Drive Cycle, NEDC)以及全球統一輕型車輛測試程序(Worldwide Harmonized Light-Duty Vehicle Test Procedure V.2, WLTP Class-2)進行模擬分析，於五次NEDC以及WLTP Class-2行車型態下，RB、ECMS、BPANN、WOA-ANN能耗分別為：50.58、39.27、47.13、39.13kWh；72.70、51.10、61.29、51.50kWh。與RB相比，ECMS、BPANN、WOA-ANN改善率分別為：22.36%、6.82%、22.64%；29.71%、15.70%、29.16%。相較於ECMS，BPANN與WOA-ANN相似度分別為：83.32、99.64%；83.37%、99.22%。本研究利用兩台快速雛型控制器，驗證ECMS理論型控制應用於車用載具之可行性，透過硬體嵌入式開發環境進行Real-time運行，在兩種行車形態下，WOA-ANN於SIL與HIL開發環境之能耗表現相似度皆高達95%，以此先行驗證應用於實車控制器之成效。This research aims to using artificial neural network (ANN) in the triple electrical power vehicle energy management strategy, for the purpose of the feasibility study of the future hardware implementation. The training set is collected by equivalent consumption minimum strategy (ECMS) that is studied in the field of hybrid vehicle energy management normally. Firstly, we construct a low-order dynamics equation that has the characteristics of the triple electrical power vehicle, including power source (a fuel cell, a lithium battery, and a supercapacitor) The key parameters were retrieved from the commercialization software Advanced Vehicle Simulator (ADVISOR). Base on the vehicle structure of the Toyota Mirai, we build the range-extended version. The power system includes a 110kW fuel cell set, a 40Ah lithium-ion battery set, and a 165F 48V supercapacitor set arrange 150kW AC motor.As the source of the ANN training set, the ECMS control strategy includes a six-stairs for-loop: the battery state-of-health (SOH), power demand the battery state-of-charge (SOC), the supercapacitor state-of-charge (SOC) the power ratio of battery to power demand(α) and the power ratio of supercapacitor to power demand(β). For the ANN training set, the input layer is set to the first four stairs for-loop and output layer is α and β. In this search, the artificial neuron network trained weights by whale optimization was compared with the traditional backpropagation method, to evaluated the benefit of the ANN accuracy, a rule based (RB) control was designed as well.Two standard test scenarios, New European Drive Cycle (NEDC) and Worldwide Harmonized Light-Duty Vehicle Test Procedure V.2 (WLTP Class-2), was chosen for the energy improvement evaluation. In five-time NEDC and WLTP Class-2, the energy consumption for RB, ECMS, BPANN and WOA-ANN was [50.58, 39.27, 47.13, 39.13kWh] and [72.70, 51.10, 61.29, 51.50kWh]. The energy improvement for ECMS, BPANN and WOA-ANN was [22.36%, 6.82%, 22.64%] and [29.71%, 15.70%, 29.16%] compared with RB, respectively. The similarity of BPANN and WOA-ANN was [83.32、99.64%]；[83.37%、99.22%] compared with ECMS. In our study, the two rapid prototype controllers were utilized to verify the feasibility of ECMS applied in vehicle hardware, the hardware-in-the-loop development environment was established to implement real-time. In NEDC and WLTP Class-2, the similarity of WOA-ANN in the SIL and HIL development environment is higher than 95%, it pre-validation the effectiveness of application in vehicle controller.