Optimization model combining sparse nonlinear dynamic identification and neural network: Research based on battery capacity prediction

With the rise of distributed power grids, efficient energy storage systems are essential to support the integration of renewable energy. Battery capacity prediction is crucial to optimize energy storage performance, extend battery life, and reduce operating costs. However, the inherent complexity of battery systems and the scarcity of high-quality data pose significant challenges to achieving accurate predictions. Traditional machine learning models, including Feedforward Neural Network (FNN), rely heavily on large datasets to achieve high accuracy. These models often fail to generalize effectively when data is limited. To address this limitation, we propose an informed machine learning model with sparse nonlinear dynamic identification (SINDy) equations as prior knowledge. This method uses cycle times as pseudo-time features and does not require strictly consistent time interval measurements, which can simplify the process of establishing battery capacity dynamic equations using SINDy. Then, the physical model obtained by SINDy is incorporated into FNN training as prior knowledge to ensure the model's prediction accuracy even when data is scarce. Experiments show that the proposed informed machine learning model has higher prediction accuracy for battery capacity with less training data.