Real-time chlorate by-product monitoring through hybrid estimation methods

Since the strict regulations regarding chlorate concentrations in drinking water and in food, there exists a need to monitor this by-product stemming from electrochlorination. Since, currently, there are no chlorate-specific sensors, Sensor Data Fusion is proposed as an alternative. The objective of this paper is to investigate and design Sensor Data Fusion algorithms that are accurate over a broader set of circumstances. Two different estimators are explored, both of which combine a first-principles model with a machine learning algorithm. The first-principles models are based on a nonlinear, reduced-order state-space model. The data-driven models investigated were multiple linear regression, K nearest neighbors, a gradient-boosting decision tree and support vector regression, with optimized hyperparameters and a two-stage validation process. It was found that the addition of a first-principles model reduced the cross-validation mean squared error by 58%, and allows accurate scaling with the fluid flow rate, when used in combination with support vector regression. Furthermore, a relatively simple hybrid approach, with state-space and data-driven models in series, was sufficient in terms of accuracy, when compared to a more complex series–parallel hybrid version. The latter does provide information regarding the free chlorine concentration and current efficiencies in real-time, as well as an estimate of the uncertainties associated with the process states. The 1 σ confidence interval converged to 14% of the chlorate estimate. The results indicate that a hybrid approach is viable in the design of a Sensor Data Fusion algorithm for chlorate monitoring, and preferable over a purely data-driven approach. • Adding features based on first-principles strongly improves a data-driven estimator. • A “virtual electron concentration” model allows applying fluid dynamics to inputs. • System states can be tracked based on a data-driven estimate, using an EKF. • Compared to data-driven methods, the CV-MSE is 58% lower, or less data is required.