| dc.description.abstract | This paper presents a chiller diagnosis system to analyze the energy consumption of the chiller and proposes the parameter adjustment strategy for chiller operation to improve its efficiency.Since energy consumption analysis involves many parameters (such as chilled water return temperature, cooling tower water temperature, etc.), the chiller model will be built for daily monitoring, fault detection, and diagnosis. In the beginning, the outliers in the raw data derived from sensors are removed, and then training and testing sets are split by stratified random sampling during data preprocessing step. This study compares the accuracy of several machine learning models and neural network. We apply Bayesian optimization and cross-validation techniques while training the model to find the best model hyperparameters quickly and impartially. The cross-validation can improve the prediction accuracy of the model since it can utilize more data to train the model. The input features of a chiller can be ranked by SHAP values according to their contribution to the model prediction after the model has been trained.In order to know whether the chiller system is operating at an efficient condition, the z-score and mean value difference between real-time efficiency and model prediction are employed as the criteria to judge the efficiency difference. In the diagnosis module, a baseline of operation parameters adjustment will be built first by KNN and X-means. After that, a greedy algorithm is applied to find the relationship between parameters and operation faults. The diagnosis module will further consider the analysis result of big data decision tree to find the efficiency difference among different chillers. The result shows that the system can automatically recommend parameter adjustment strategies for different chillers after a simple initial setting. The system allows users to modify many options, such as modeling options and SQL connection settings. Also, users can use the GUI to check the efficiency analysis results, features influence, recommend parameters adjustment strategy, and expected efficiency curve before and after optimization. Furthermore, users can retrain the model based on the analysis methods they want by using different period data and desired parameter settings. | en_US |