The issue is that the predict() method does not apply the same preprocessing as the fit() method.
During training, the model uses preprocessed data (via self.preprocess_input()), but during prediction, it directly uses raw input. This mismatch causes prediction errors because the model expects preprocessed input.
By updating the predict() method to call self.preprocess_input(model_input.copy()), both training and prediction use consistent feature transformations, allowing predictions to be computed successfully.