Increasing power plant efficiency with clustering methods and Variable Importance Index assessment

Power plant performance can decrease along with its life span, and move away from the design and commissioning targets. Maintenance issues, operational practices, market restrictions, and financial objectives may lead to that behavior, and the knowledge of appropriate actions could support the system to retake its original operational performance. This paper applies unsupervised machine learning techniques to identify operating patterns based on the power plant’s historical data which leads to the identification of appropriate steam generator efficiency conditions. The selected operational variables are evaluated in respect to their impact on the system performance, quantified by the Variable Importance Index. That metric is proposed to identify the variables among a much wide set of monitored data whose variation impacts the overall power plant operation, and should be controlled with more attention. Principal Component Analysis (PCA) and k-means++ clustering techniques are used to identify suitable operational conditions from a one-year-long data set with 27 recorded variables from a steam generator of a 360MW thermal power plant. The adequate number of clusters is identified by the average Silhouette coefficient and the Variable Importance Index sorts nine variables as the most relevant ones, to finally group recommended settings to achieve the target conditions. Results show performance gains in respect to the average historical values of 73.5% and the lowest efficiency condition records of 68%, to the target steam generator efficiency of 76%.