Artificial neural networks for long-term electric load modeling and forecasting

Farqad AlKhal, Purdue University

Abstract

Artificial neural networks offer a data modeling tool that can prove to be powerful for many real world applications. Modeling and predicting demand for energy is one application that can potentially benefit from that tool. However, whenever neural networks are used, issues pertaining to model construction given limited data, parameter estimation, and assessment of model performance have to be dealt with. Estimating neural network weights using observed examples, called supervised learning can be formulated as a nonlinear minimization problem. The objective function in this problem has been shown to suffer from local minima and replicas of the global minimum. Hence, the problem of finding the “best” values for the network weights becomes a nonlinear multiexternal minimization problem. To solve this problem, local search techniques like gradient descent methods are not “very” computationally demanding, but are myopic in nature and can get trapped in local minima. On the other hand, random search techniques like simulated annealing have a built in mechanism that enables them to escape local minima, but they require a large number of function evaluations. We propose a solution method that uses a gradient search technique on smoother transformed versions of the objective function, make use of multivariate Gaussian quadrature to approximate the transformation, and suggest an algorithm that uses the Quasi-Newton method to minimize those transformations and traces their solutions as the amount of smoothing is gradually reduced to zero. Standard test problems from the nonconvex optimization literature and synthetic neural network problems of various sizes are used to evaluate the performance of the proposed algorithm and compare it to that of a local search, and a stochastic global search technique. Furthermore, the problem of overlearning in neural networks is addressed, and existing methods to cure it are reviewed. Finally, a case study using real data is conducted to model and predict quarterly demand for electricity in a given residential service area. The study illustrates the potential benefit gained from using neural networks as a highly flexible nonlinear mapping tool which can be utilized for both predictive and explanatory purposes.

Degree

Ph.D.

Advisors

Sparrow, Purdue University.

Subject Area

Operations research|Industrial engineering|Artificial intelligence

COinS