Pedro M. Ferreira

Genetic assisted selection of RBF model structures for greenhouse inside air temperature prediction

This paper presents results on the application of Multi-Objective Genetic Algorithms to the selection of Radial Basis Function Neural Networks structures. The neural networks are to be incorporated in a real-time predictive greenhouse environmental control strategy, as predictors of the inside air temperature. Previous research conducted by the authors modelled the inside air temperature as a function of the inside relative humidity and of the outside temperature and solar radiation. A second-order model structure previously selected in the context of dynamic temperature models identification was used. Several training and learning methods were compared, and the application of the Levenberg-Marquardt optimisation method was found to be the best way to determine the neural network parameters. The application of correlation-based model-validity tests revealed that the validity of such a second-order model structure could be manually improved after inspection of the tests results. Both network performance and validity are certainly affected by the number of neurons, the input variables considered and the time delays used. As the number of alternatives is huge, Multi-Objective Genetic Algorithms are applied here to the selection of network inputs and number of neurons.

Online Sliding-Window Methods for Process Model Adaptation

Online learning algorithms are needed when the process to be modeled is time varying or when it is impossible to obtain offline data that cover the whole operating region. To minimize the problems of parameter shadowing and interference, sliding-window-based algorithms are used. It is shown that, by using a sliding-window policy that enforces the novelty of the data it stores and by using a procedure to prevent unnecessary parameter updates, the performance achieved is improved over a first-in-first-out (FIFO) policy with fixed interval parameter updates. Important savings in computational effort are also obtained.

EVOLUTIONARY MULTIOBJECTIVE DESIGN OF RADIAL BASIS FUNCTION NETWORKS FOR GREENHOUSE ENVIRONMENTAL CONTROL

Abstract In this work a multiobjective genetic algorithm is applied to the identification of radial basis function neural network coupled models of humidity and temperature in a greenhouse. Models are built as one-step-ahead predictors and then used iteratively to produce long term predictions. The number of neurons and input terms used in both models define the search space. Two combinations of performance and complexity criteria are used to steer the selection of model structures, resulting in distinct sets of solutions. It is shown that minimisation of one-step-ahead prediction errors negatively influences long term prediction performance. Long term prediction results are presented for a pair of models selected from sets of models obtained in the experiments.

Evolutionary Multiobjective Neural Network Models Identification: Evolving Task-Optimised Models

In the system identification context, neural networks are black-box models, meaning that both their parameters and structure need to be determined from data. Their identification is often done iteratively in an ad-hoc fashion focusing the first aspect. Frequently the selection of inputs, model structure, and model order are underlooked subjects by practitioners, because the number of possibilities is commonly huge, thus leaving the designer at the hands of the curse of dimensionality. Moreover, the design criteria may include multiple conflicting objectives, which gives to the model identification problem a multiobjective combinatorial optimisation character. Evolutionary multiobjective optimisation algorithms are particularly well suited to address this problem because they can evolve optimised model structures that meet pre-specified design criteria in acceptable computing time. In this article the subject is reviewed, the authors present their approach to the problem in the context of identifying neural network models for time-series prediction and for classification purposes, and two application case studies are described, one in each of these fields.

Neural network PMV estimation for model-based predictive control of HVAC systems

Heating, Ventilating and Air Conditioning (HVAC) systems are used to provide adequate comfort to occupants of spaces within buildings. One important aspect of comfort, the thermal sensation, is commonly assessed by computation of the Predicted Mean Vote (PMV) index. Model-based predictive control may be applied to HVAC systems in existing buildings in order to provide a desired degree of thermal comfort and simultaneously achieve significant energy savings. This control strategy may be formulated as a discrete optimisation problem and solved by means of structured search techniques. Finding the optimal solution depends on the ability of computing many PMV values in a small amount of time. As the PMV formulation involves iterative computations consuming variable time, it is crucial to have a method for fast, possibly constant execution time, computation of the PMV index. In this paper it is experimentally shown that an Artificial Neural Network (ANN) can estimate the PMV index with varying degrees of efficiency over the trade-off of accuracy versus computational speed-up.

Energy savings in HVAC systems using discrete model-based predictive control

The paper addresses the problem of controlling an heating ventilating and air conditioning system with the purpose of achieving a desired thermal comfort level and energy savings. The formulation uses the thermal comfort as a restriction and minimises the energy spent to comply with it. This results in the maintenance of thermal comfort and on the minimisation of energy, which in most operating conditions are conflicting goals requiring some sort of optimisation method to find appropriate solutions over time. In this work a discrete model based predictive control methodology is applied to the problem. It consists of three major components: the predictive models, implemented by radial basis function neural networks identified by means of a multi-objective genetic algorithm; the cost function that will be optimised to minimise energy consumption and provide adequate thermal comfort; and finally the optimisation method, in this case a discrete branch and bound approach. Each component will be described, and experimental results obtained within a classroom will be presented, demonstrating the feasibility and performance of the approach. Finally the energy savings resulting from the application of the method are estimated.

Solar radiation prediction using RBF Neural Networks and cloudiness indices

In this paper, Artificial Neural Networks are applied to multi-step long term solar radiation prediction. The networks are trained as one-step-ahead predictors and iterated over time to obtain multi-step longer term predictions. Auto-regressive and Auto-regressive with exogenous inputs solar radiation models are compared, considering cloudiness indices as inputs in the latter case. These indices are obtained through pixel classification of ground-to-sky images. The input-output structure of the neural network models is selected using evolutionary computation methods.

Application of computational intelligence methods to greenhouse environmental modelling

In order to implement a model-based predictive control methodology for a research greenhouse several predictive models are required. This paper presents the modelling framework and results about the models that were identified. RBF neural networks are used as non-linear auto-regressive and non-linear auto-regressive with exogenous inputs models. The networks parameters are determined using the Levenberg-Marquardt optimisation method and their structure is selected by means of multi-objective genetic algorithms. By network structure we refer to the number of neurons of the networks, the input variables and for each variable considered its lagged input terms. Two types of models were identified: process models (greenhouse climate) and external disturbances (external weather). Pseudo-random binary signals were employed to generate control input commands for the greenhouse actuators, in order to build input/output data sets suitable for the process models identification. The final model arrangement consists of four interconnected models, two of which are coupled, providing greenhouse climate and external weather long term predictions.

Discrete Model Based Greenhouse Environmental Control Using the Branch & Bound Algorithm

Abstract In this paper we propose the application of the Branch-and-Bound search algorithm to discrete model-based predictive control of greenhouses. The temperature control strategy is a mixture of temperature integration and difference between day and night temperatures. The general approach is presented and strategies are proposed in order to achieve a faster coverage of the solutions search space with reduced probability of loosing the optimal solution. The control energy requirements depend largely on the cost function coefficients and the evolution of the external climate. Fixed coefficients do not fully exploit the external climate predicted evolution in order to reduce energy consumption. A simple method is proposed to adapt on-line the cost function coefficients in a way that reduces energy consumption without significantly affecting control accuracy. The methods are briefly described and a subset of experimental and simulation results are presented.

Model Based Predictive Control of HVAC Systems for Human Thermal Comfort and Energy Consumption Minimisation

Abstract The problem of controlling a heating ventilating and air conditioning system in a single zone of a building is addressed. Its formulation is done in order to maintain acceptable thermal comfort for the occupants and to spend the least possible energy to achieve that. In most operating conditions these are conflicting goals, which require some sort of optimisation method to find appropriate solutions over time. In this work a model based predictive control methodology is proposed. It consists of three major components: the predictive models, implemented by radial basis function neural networks identified by means of a multi-objective genetic algorithm; the cost function that will be optimised to minimise energy consumption and provide adequate thermal comfort; and finally the optimisation method, in this case a discrete branch and bound approach. Each component will be described, and experimental results obtained within a classroom will be presented, demonstrating the feasibility and performance of the approach.