Research on power prediction model for waste heat recovery systems based on ESO feature extraction

When a heavy-duty internal combustion engine adopts electric turbo-compounding for waste heat recovery, forming a hybrid waste heat recovery system, predicting its output power becomes extremely difficult due to the system’s strong nonlinearity, multi-timescale dynamic characteristics, and complex coupling relationships. This paper proposes a hybrid model combining an Extended State Observer (ESO) with Optuna-tuned hyperparameters and a Multilayer Perceptron (MLP). The ESO treats unknown internal dynamics and external interference as a single total disturbance and estimates it in real time. Optuna automatically finds the best observer gain settings, allowing the model to extract physically meaningful dynamic features from raw data. These features then combine with historical operating data to build an input matrix for the MLP, which learns the complex nonlinear relationships to predict output power accurately. Simulation tests used engine fuel flow rates and total output power data from four standard driving cycles: IMO240, Highway, HUDDS, and ARTEMIS. Results show the proposed model outperforms mainstream baselines like Transformer, LSTM, and EfficientKAN on accuracy metrics including RMSE, MAE, and R ² . Evaluation using Standard Deviation of Error (SDE) confirms its superior stability when handling data fluctuations under complex operating conditions, offering reliable support for developing energy management strategies in hybrid power systems.