Hybridization of vehicles has been proven a good way to reduce fuel consumption significantly. Working prototypes of a series hydraulic hybrid vehicle (SHHV) are already under testing. The power split strategy for those prototypes is a rule-based controller, or called a “bang-bang” controller. The controller is designed based on engineer's intuition, to keep the engine working in the region with high efficiency and low fuel consumption rate. One of the problems of that design is that it only takes one component of the hydraulic hybrid system, the internal combustion engine, into account. It is a device centered rather than system centered design. As a result, the potential of the hydraulic hybrid system is not fully realized.
A more efficient power split strategy is conducted based on the Deterministic Dynamic Programming (DDP), which has been proved a powerful tool for optimal control. However, the DDP is a looking-forward tool, which means it uses the future driving conditions to split the power between the two sources for optimization. Successful applications of DDP used standard driving cycles as the known driving conditions. However, DDP is not applicable where the driving cycle is unknown. This means that the DDP could not be applied in real-time, unless the future driving conditions could be found.
The driving conditions in our everyday commute are extremely different with the typical driving cycles. And different drivers have different driving habits. However, a specific driver has a certain “driving cycle” for a certain commute, although which is not a standard one. As long as the certain “driving cycle” is known, The DDP algorithm could be applied for optimization. Artificial neural network (ANN) has the ability to “learn” the “driving cycle” from a certain driver and then to “predict” the driving conditions before its happening. The “prediction” method is the “time-series forecasting” method. ANN is a good tool for time series forecasting and has also been shown a better way for long term prediction. The ANN is conducted using the software MATLAB/Simulink. A three-layer feed-forward static ANN is built up in the Simulink environment.
The ANN model was able to predict the driving conditions with a twenty seconds window size which has been proven a tradeoff between the forecasting accuracy and the time consumed. The error between the predicted value and the desired value is within an accepted range. The network is tested based on three different driving cycles: federal urban driving schedule, city urban dynamometer driving schedule and highway urban dynamometer driving schedule respectively.