Energy optimization in the process industry using AI

Important results from the project

The aim of the project was to use AI technology to investigate the possibilities of using the energy within the process in a more efficient way, thus reducing the need of fossil fuel as additional energy. The project has shown that an AI-based model, which is built and trained correctly, can find a connection between the various sub-processes and thus can be used to forecast all states. An optimization of the process control shows that an alternative way of controlling the recovery boiler can minimize the need for fossil supplement fuel.

Expected long term effects

The model´s accuracy proved to be high for most subprocesses during verification. Given only the production planning as input, the model can forecast all tower levels, with a slight deviation, 50 hours ahead. The use of oil has a somewhat poorer accuracy. On examination, it was found that the use of oil may be due to causes other than a lack of steam alone, causes not found in the training data. The optimization of the process control has resulted in that all the oil that the model can predict can be minimized by an alternative control of mainly the soda boiler.

Approach and implementation

The model was trained on process data from about 25 different measurement points and a total of 57,000 measurements. The model is a black box model and the information about how the sub-processes are connected is only obtained through the training data. The optimization was carried out with reinforcement learning, where, among other things, oil use was set as a negative event. The optimization was carried out over a total of 1900 hours with the actual limitations of the process as limits. The project has demonstrated great potential for using AI technology for optimization of a paper mill.