Loading...
1 results
Search Results
Now showing 1 - 1 of 1
- Artificial neural network models: data selection and online adaptationPublication . Khosravani, Hamid Reza; Ruano, A. E.; Ferreira, Pedro Miguel Frazão F.Energy consumption has been increasing steadily due to globalization and industrialization. Studies have shown that buildings have the biggest proportion in energy consumption; for example in European Union countries, energy consumption in buildings represents around 40% of the total energy consumption. Hence this PhD was intended towards managing the energy consumed by Heating, Ventilating and Air Conditioning (HVAC) systems in buildings benefiting from Model Predictive Control (MPC) technique. To achieve this goal, artificial intelligence models such as neural networks and Support Vector Machines (SVM) have been proposed because of their high potential capabilities of performing accurate nonlinear mappings between inputs and outputs in real environments which are not noise-free. In this PhD, Radial Basis Function Neural Networks (RBFNN) as a promising class of Artificial Neural Networks (ANN) were considered to model a sequence of time series processes where the RBFNN models were built using Multi Objective Genetic Algorithm (MOGA) as a design platform. Regarding the design of such models, two main challenges were tackled; data selection and model adaptation. Since RBFNNs are data driven models, the performance of such models relies, to a good extent, on selecting proper data throughout the design phase, covering the whole input-output range in which they will be employed. The convex hull algorithms can be applied as methods for data selection; however the use of conventional implementations of these methods in high dimensions, due to their high complexity, is not feasible. As the first phase of this PhD, a new randomized approximation convex hull algorithm called ApproxHull was proposed for high dimensions so that it can be used in an acceptable execution time, and with low memory requirements. Simulation results showed that applying ApproxHull as a filter data selection method (i.e., unsupervised data selection method) could improve the performance of the classification and regression models, in comparison with random data selection method. In addition, ApproxHull was employed in real applications in terms of three case studies. The first two were in association with applying predictive models for energy saving. The last case study was related to segmentation of lesion areas in brain Computed Tomography (CT) images. The evaluation results showed that applying ApproxHull in MOGA could result in models with an acceptable level of accuracy. Specifically, the results obtained from the third case study demonstrated that ApproxHull is capable of being applied on large size data sets in high dimensions. Besides the random selection method, it was also compared with an entropy based unsupervised data selection method and a hybrid method involving ApproxHull and the entropy based method. Based on the simulation results, for most cases, ApproxHull and the hybrid method achieved a better performance than the others. In the second phase of this PhD, a new convex-hull-based sliding window online adaptation method was proposed. The goal was to update the offline predictive RBFNN models used in HVAC MPC technique, where these models are applied to processes in which the data input-output range changes over time. The idea behind the proposed method is capturing a new arriving point at each time instant which reflects a new range of data by comparing the point with current convex hull presented via ApproxHull. In this situation the underlying model’s parameters are updated based on the new point and a sliding window of some past points. The simulation results showed that not only the proposed method could efficiently update the model while a good level of accuracy is kept but also it was comparable with other methods.