Application of Reinforcement Learning to Batch Distillation

An important amount of work exists on the topic of optimal operation and control of batch distillation though it is still based on the assumption of an accurate process model being available. While this assumption is valid from a theoretical point of view, there will always remain the challenge of practical applications. Reinforcement Learning (RL) has been recognised already as a particularly suitable framework for optimizing batch process operation however no application to batch distillation has been reported. Thus, this paper presents RL as an automatic learning approach to batch distillation. The methodology is exemplified using various case studies. INTRODUCTION Distillation is one of the most widely used unit operations in the fine chemical, petroleum and pharmaceutical industries. It is one of the oldest methods of separation of liquid mixtures into their various components depending on differences in boiling points of liquids and relative volatility. The rising importance of high-value-added, lowvolume specialty chemicals has resulted in a renewed interest in batch processing technologies (Diewkar, 1995) and the drive for optimum operation is ever present. Batch distillation is an important and widely used separation process in batch process industry. Its main advantage over continuous operation is the ability to be used as a multi-purpose operation for separating mixtures into their pure components using a single column. Batch distillation can also handle a wide range of feed compositions with varying degrees of difficulty of separation (e.g. wide ranges of relative volatilities and product purities). Although the typical consumption of energy is more than in continuous distillation, more flexibility is provided with less capital investment (Luyben, 1992). However, besides the flexibility in the operation of batch distillation columns, a range of challenging design and operational problems occur due to its inherent unsteady state nature. LITERATURE SURVEY The main sequence of events in operating a batch distillation column starts with the feed charged into the reboiler. The column is then operated at total reflux until the column reaches steady state. This initial phase is known as the start-up phase. In the second phase, or production phase, light component product is collected into a product tank until its average composition drops below a certain specified value. This cut is referred to as the main cut (The 1 st main cut is sometimes preceded by taking off the low boiling impurities at a high reflux ratio). After that, the first intermediate distillate fraction (off-cut or slop cut) is produced and stored in a different tank. This procedure is repeated with a second main cut and second slop cut and so on until the concentration of the heaviest component, in the reboiler of the column, reaches a specified value. At the end of the batch, the operation of the distillation column goes through a shutdown phase. Slop cuts contain the material distilled, which does not meet specification. Considerable work in slop handling strategies has been reported in the literature ((Bonny et. al., 1994) and (Mujtaba and Macchietto, 1992)). On the other hand, a totally different operating policy is the cyclic operation of a batch distillation column. In the case of a regular column, the cyclic operation could be characterised by repeating a three period operation (Sorensen, 1997): Filling, Total Reflux, and Dumping. The main manipulated variable, in the process of controlling a batch distillation column, is the reflux ratio. The frequently used and conventional approach towards controlling the operation of a batch distillation column, during the production of main cuts, is either to operate at constant reflux ratio or to operate at a varying reflux ratio (constant distillate composition). During operation at constant reflux ratio, the distillate composition is allowed to vary resulting in a simpler strategy and hence it is more commonly used in industry. The second approach is conducted by maintaining a fixed overhead composition while varying the reflux ratio. The two approaches used are simple but provide sub-optimal results.