Predicting Fouling Progress in Water Cooled Chillers Using Bidirectional LSTM Neural Networks and Synthetic Data

The primary challenge in developing a practical predictive model for equipment cleaning is determining the variables the neural network should calculate, focusing on a cost-benefit analysis. Given the impracticality of using real installation data due to the slow development of deposits, synthetic data were generated using a theoretical model in Engineering Equation Solver, considering variations in external wet bulb temperature and chiller parameters. These simulations, improving upon previous studies, trained neural networks to recognize fouling patterns in new data. The final model, suitable for real applications, was created using Python and executed in Google Colab Notebook, leveraging TensorFlow, Keras, NumPy, Pandas, and scikit-learn libraries. A Bidirectional Long Short-Term Memory (LSTM) architecture with the "tanh" activation function and Adam optimization was adopted, balancing accurate modeling and avoiding overfitting. The model was trained with 10,000 data points, using time, wet bulb temperature, condensation temperature, conductance, pump pressure, and tube length as inputs, and Rf as the output. The results highlighted fouling progression and its impact on chiller performance, demonstrating the viability of this approach. Future work includes testing in real installations, adapting the model using transfer learning, and exploring different architectures and additional physical parameters.