Combined neural networks and predictive control for heat exchanger networks operation

Abstract

Optimal operation of integrated heat exchangers is a challenging task in the field of process control due to system nonlinearities, disturbances and adequate model identification. This paper describes the design of an advanced neural network predictive control (NNPC) applied to a heat exchanger network. A case study with two hot and one cold streams, through three counter-current heat exchangers is used to test the proposed strategy. A lumped dynamic model is built based on the concept of multi-cells topology (mixed tanks), where the hot and cold cells are connected by a wall element throughout the heat exchanger length. Each cell is assumed perfectly mixed and all physical properties are constant. A distributed behavior is achieved by increasing the number of cells. The main assumptions of the lumped model are constant temperature in each cell, heat exchanger volume and area equally distributed between cells and negligible heat loss to the environment. The predictive controller relies on a neural-based model of the plant that is used to identify the system and to predict future performance over a predefined horizon. Results were compared to a traditional controller, and the control performance was improved when compared to the Ziegler-Nichols tuning method.