Control and optimisation of coagulant dosing in drinking water treatment

Abstract

Correct coagulant dosage is necessary for the efficient operation of conventional drinking water treatment plants, yet no accurate or automated way of determining this exists. Streaming current (SC) is a measurement of charge on particles in water and is useful in feedback control of coagulant dosage. Analysis of the movement of change within a SC sensor can provide some explanation of its slow response, while signal processing utilising Fourier analysis improves the instrument's bandwidth. Presently inaccurate manual jar tests are the only way determine the SC required for best coagulation. An online automated jar tester is presented to improve on this. It uses an automatic sampling system that takes a sample from the process stream. An optimisation algorithm makes repeated step adjustments to the SC set point and gradually moves it in the direction of improving jar test results. The system was evaluated on both a small-scale model and a full-scale plant. Noise in the test measurements means the optimal set point cannot be located accurately enough, but the results indicate that this is possible. Greater accuracy would allow optimisation of turbidity and costs for multiple chemicals. A representative neural network model can be made of the dynamic relationship between coagulant dosage and streaming current in a scale model, with an alkali dosed to simulate a disturbance. In a rapid mixer, the measured response is significantly slower than the true response. Several common types of linear controller are designed and their performance at set point tracking and disturbance rejection is compared on this system. Model predictive control with a Kalman filter performs best in these tests, while the self-tuning regulator has benefits when the rate of set point change is slower. A non-linear feed-forward radial basis function network that adapts to the system's steady-state inverse can effectively augment a linear controller for this system. Adaptation rules based on vector eligibility are derived from dynamic back-propagation and extended to the general dynamic non-linear case. This can result in a useful and efficient feed-forward neural controller for dosing systems that can be represented by a Wiener model.