以實車數據建構電動車剩餘里程之深度學習預測模型

Abstract

近年來，隨著油價上漲和環保意識抬頭，電動車的市場在台灣蓬勃發展。相較於傳統油車，電動車具備更高的能源效率並且對於環境的污染相對較少。然而國人對於電動車的購買以及使用依然有些許的隱憂，其中包含充電樁設立數量不足、對於電池電量消耗的不熟悉、過長的充電時間以及行駛里程上的限制，而上述的問題都會造成電動車駕駛者對於剩餘里程的焦慮，不確定剩餘電量是否能到達目的地。本研究之目標在於發展深度學習的模型來預測電動車剩餘里程，以得到更精準的剩餘里程預估來緩解駕駛者的焦慮。 本研究所採用的預測模型包括支持向量回歸 (Support vector regression, SVR)、人工神經網路 (Artificial neural network, ANN)、遞迴神經網路 (Recurrent neural network, RNN) 以及其延伸長短期記憶網路 (Long short-term memory, LSTM)。預測模型的輸入資料主要參考車輛物理模型以及輔助系統耗電，包括剩餘電量、車速、坡度、加速度、風速、天氣、空調系統的設定、車燈開啟狀態等，預測模型之輸出則為單位時間耗電量，再結合車速可得到單位耗電量可行駛的里程，以剩餘電量除以單位電量里程 (unit energy mileage)，即可預估剩餘里程。 前述模型以華創車電提供之5台LUXGEN S3 EV電動車在夏季的行駛資料進行訓練，結果顯示，各模型對單位電耗里程之預測以2-time-step LSTM 模型預測結果最佳，其餘依次為1-time-step LSTM、2-time-step RNN、1-time-step RNN、ANN，SVR模型殿後，平均絕對百分比誤差分別為12.3%、15.7%、27.9%、31.3%、35.7%以及41.4%，顯見以2-time-step LSTM 模型預測單位耗電里程最具可行性。 本研究亦對2-time-step LSTM 進行參數分析，發現：1. 當車速在20至100 km/h之間，可行駛里程隨車速遞增; 2. 當剩餘電量遞減，單位電量里程隨之遞增; 3. 當車輛加速度變化越劇烈，可行駛里程遞減。 最後，本研究提出兩種向駕駛者提供電池用電情況的情境。第一種情境是將預估之剩餘里程即時通知駕駛者，但為了避免預估之剩餘里程隨駕駛情況變化而上下振盪，可對預估里程取移動平均。第二種情境適用於駕駛者已有既定行程，依據旅程路徑及當時情況，可利用前述模型預測整趟旅程所需的耗電量。這兩種資訊應有助於駕駛者了解電動車狀況，並有助於緩解其焦慮。

With the rise of global awareness in environmental protection and the gradually increasing oil prices, the automobile industry of electric vehicle (EV) is growing rapidly. Compare with conventional fuel vehicles, EVs are not only more environment-friendly but also more energy-efficient and therefore are gaining ever amount of attention in many countries. However, there are still a few obstacles to EVs that restrict the development of EVs in Taiwan currently. In general, EV drivers in Taiwan are more concerned about potential interruption caused by battery depletion, owing to sparsely distributed charging stations, long charging time, and limited mileage. In contrast to conventional vehicle drivers, EV drivers have “range anxiety” that they worry about the feasibility of reaching the destination. The objective of this study is to develop a model to predict the remaining range of an EV, hoping that the driver’s anxiety may be alleviated by such information. The estimation models studied in this research include support vector regression (SVR), artificial neural network (ANN), recurrent neural network (RNN), and long short-term memory (LSTM). The input data, selected based on the physical model of the vehicle, comprises state of charge (SOC), speed, acceleration, the slope of the road, wind speed, light, weather condition, the setting of the auxiliary system, etc. The output of the model is the energy consumption in a time step, which can be used to calculate the unit energy mileage and further to estimate the remaining driving range at current SOC. The models were trained using the data collected from five LUXGEN S3 EVs in the summer. It is found that the two-time-step LSTM has the best performance, followed by the single-time-step LSTM, the two-time step RNN, the single-time step RNN, ANN, and SVR comes last. The mean absolute percentage errors of these models are 12.3%, 15.7%, 27.9%, 31.3%, 35.7%, and 41.4%, respectively. It is seen that the two-time-step LSTM is most feasible for the prediction of specific energy consumption. Parametric analysis was conducted on the two-time-step LSTM model, and two findings were drawn: 1. the total mileage increases with the vehicle speed within the speed range of 20 to 100 km/h; 2. the unit energy mileage increases as the value of SOC decreases; 3. The total mileage increases as the acceleration changes drastically. Two scenarios were proposed as to how the battery condition is presented to the driver. The first scenario is to estimate the real-time remaining range then present it to the driver. Taking a moving average may help smooth out the short-term fluctuation of the estimated range. The other scenario applies when there is a route plan. The route and traffic conditions can be used to predict the energy consumption of the whole trip. Both types of information should be able to help the driver better grasp the vehicle condition and relieve anxiety as driving an electric vehicle.