Abstract
Modeling is finding the inherent relationship between process input and output. Generally a mathematical expression is developed based on process understanding. This classical method is termed phenomenological modeling approach. It is time consuming as it requires detailed understanding of the process along with related and associated phenomena along with rigorous mathematical understanding. Present days with increase in computation facilities, we can opt for an empirical input – output mapping, based on historic process data. No in depth knowledge of the process is essential. Artificial Neural Networks (ANNs) are very popular now-a-days for various engineering and scientific modeling purposes as for most of the recent day processes we have historic process data stored or recorded. Another advantage is that ANNs have ability to obtain multi-input multi-output relationship directly. Process inherent nonlinearity makes the conventional modeling task even more difficult due to varying amount of nonlinear interactions involved in between different process variables. ANNs can handle even nonlinear processes and even tolerate some amount of noise present in the process data.
Process optimization is a mathematical technique followed to dig out the best possible outcome among all the feasible solutions of a process. Often we are very eager to maximize profit of any venture and quite naturally try to minimize by product formation. So there are several examples where we can apply the optimization technique to find the best suited path or best combination of variables in order to achieve the desired best result. Conventional optimization techniques are gradient based and hence computation intensive as they require computation of first and second order derivatives. These methodologies cannot be operated if the objective function has discontinuity in it or the gradient computation is almost impossible due to complexity. There are new generation optimization tools like Genetic Algorithms (GAs) and Differential Evolution (DE) which are evolved based on evolution process. They require computation of functional value only, not the 1st and / or higher order derivatives. Again they are population based so with the number of iterations they gradually update and converge towards the optimal solution. Another inherent advantage of these methods are that they can able to pick the global optimum out of several local optima. Even they can be applied for multi-objective paerto optimizations as well.
The lecture will introduce ANNs as an efficient modeling tool and discuss the efficacies of GA and DE for optimization. A case study of a petrochemical company for their pilot plant studies will also be discussed.