Machine Learning Part 2: The Model and Leveraging the Results
Part 2 – The Model and Leveraging the Results – A Detailed Perspective of Steps for Machine Learning and Predictive Chilled Water System Analytics in Facility Applications
ZLATKO VASILKOSKI, CHIEF SCIENTIST, FACILITYCONNEX
If you have not already, check out Part 1 here
Creating Machine Learning Models
THE DATASET
The dependent derived chiller KPIs and their hourly, daily and seasonal changes are shown in Figure 2.
Figure 2. Dependent KPIs for the chiller model are: OAT – Outside air temperature (blue), CHW_DeltaT – Chilled water temperature difference (green), CDW_DeltaT- Condenser water temperature difference (red). The time period shown is between June 2017 and Aug 2018. Daily and seasonal changes in the data can be observed, giving insight into which machine learning model needs to be used. It can be noticed that when the chiller status is on, the CHW_DeltaT is larger than CDW DeltaT. This is also the case for January 8, 2018 when the OAT is -12F. The drops in the red and the green curves (DeltaT’s) are due to the chiller status being off.
This figure shows the historical data period between June 2017 and Aug 2018, used to create the chiller model. As described previously in the preprocessing, only certain values of this data were used in creating the chiller model, but these values span the hourly, daily and seasonal changes which is important for the robustness of the created model.
The outside air temperature and the independent derived chiller KPI, the electric demand per ton, are shown in Figure 3.
Figure 3. The outside air temperature (blue) OAT is measured on the right vertical axis, the electric demand (green) in kW/Ton is on the left vertical axis. The seasonal and the daily changes can be noticed as well as the periods when the status of the chiller is off (electric demand being zero). It also can be noticed that as the OAT is increased during the summer months, the electric demand is also increased.
THE MODELS
The next step is to create various regression models for the equipment (the chiller), compare their performance and select the best one. In developing the equipment model for a first time, a good data science practice is to start with the simplest regression model, the multivariate linear regression and see if it gives satisfactory predictions on the validation data. If not, then progress to creating more complicated models using other machine learning techniques including neural networks.
For this purpose, to create the regression model, we have used the normal equation implemented in TensorFlow. Alternatively, we have created a recurrent neural network model, typically used for time series data, using Keras libraries. It should be noted that the dependence of the electric demand per ton on the given dependent variables is relatively simple and the results from both models were comparable. Including the humidity and the dewpoint will introduce much larger complexity in the model and justify the neural network approach.
The model predictions are displayed in Figure 4. It can be seen in July 2018 that the chiller’s efficiency began to degrade. The model, based on the historical data, provides a baseline for this particular chiller.
Figure 4. Chiller model results for the period from April to August 2018. The actual electric demand in kW/Ton is in blue while the optimal (baseline) electric demand, as predicted by the chiller model is in green. It can be seen that the model captures the seasonal, daily and hourly variations. The actual demand going away from the predicted baseline, starting in July is an indication of poor chiller performance.
This predictive model can then be used for predictive maintenance, by diagnosing poor chiller performance for which we need to identify the corresponding events. In general, manually corelating events and causes with these performance models can be a cumbersome process that requires a lot of analysis. But by integrating both the predictive analytics results with the prescriptive analytics results in a single view using the FCX Voice UI, this process is streamlined. This combination offers the users and easy way to find the reasons (prescriptive FDD results) of why the equipment is not performing efficiently (predictive results). This is illustrated in Figure 5.
Figure 5. Prescriptive and predictive results shown in a single view using the FCX Voice UI. The bottom part of the figure shows the results of the predictive analytics. The actual electric demand in kW/Ton is in blue while the optimal (baseline) electric demand, as predicted by the chiller model is in green. The upper part of the figure shows the results of the prescriptive fault detection (FDD) analytics with the 13 different applied analytics to this equipment (chiller).
From the model, additional regions of inefficiency and the reasons for them can be identified. This is illustrated in Figure 6.
Figure 6. Unusually high spike in Leaving Chilled Water Temperature (green line) between March 02-13, 2018, and the relation to the FDD Analytics Findings.
CONCLUSION
The FCX predictive methodology, based on variety of machine learning methods and equipment domain knowledge, is capable of detecting events that mandate the need for maintenance. The accumulation of more and more customer data just improves these models. In addition, collecting operational and maintenance data are also used to develop advanced analytics approaches for modeling the equipment.