Percorrer por autor "Pestana, Rui"
A mostrar 1 - 4 de 4
Resultados por página
Opções de ordenação
- Evolving RBF predictive models to forecast the Portuguese electricity consumptionPublication . Ferreira, P. M.; Ruano, Antonio; Pestana, Rui; Kóczy, László T.The Portuguese power grid company wants to improve the accuracy of the electricity load demand (ELD) forecast within an horizon of 24 to 48 hours, in order to identify the need of reserves to be allocated in the Iberian Market. In this work we present some preliminary results about the identi cation of radial basis function (RBF) neural network (NN) ELD predictive models and about the performance of a model selection algorithm. The methodology follows the principles already employed by the authors in di erent applications: the NN models are trained by the Levenberg-Marquardt algorithm using a modi ed training criterion, and the model structure (number of neurons and input terms) is evolved using a Multi-Objective Genetic Algorithm (MOGA). The set of goals and objectives used in the MOGA model optimisation reflect different requirements in the design: obtaining good generalisation ability, good balance between one-step-ahead prediction accuracy and model complexity, and good multi-step prediction accuracy. A number of experiments were carried out, whose results are presented, producing already a number of models whose predictive performance is satisfactory.
- Forecasting the Portuguese electricity consumption using least-squares support vector machinesPublication . Ferreira, P. M.; Cuambe, D. I.; Ruano, Antonio; Pestana, RuiThe subject of this paper is the multi-step prediction of the Portuguese electricity consumption profile up to a 48-hour prediction horizon. In previous work on this subject, the authors have identified a radial basis function neural network one-step-ahead predictive model, which provides very good prediction accuracy and is currently in use at the Portuguese power-grid company. As the model is a static mapping employing external dynamics and the electricity consumption time-series trend and dynamics are varying with time, further work was carried out in order to test model resetting techniques as a means to update the model over time. In on-line operation, when performing the model reset procedure, it is not possible to employ the same model selection criteria as used in the MOGA identification, which results in the possibility of degraded model performance after the reset operation. This work aims to overcome that undesirable behaviour by means of least-squares support vector machines. Results are presented on the identification of such model by selecting appropriate regression window size and regressor dimension, and on the optimization of the model hyper-parameters. A strategy to update this model over time is also tested and its performance compared to that of the existing neural model. A method to initialize the hyper-parameters is proposed which avoids employing multiple random initialization trials or grid search procedures, and achieves performance above average.
- Improving the identification of RBF predictive models to forecast the Portuguese electricity consumptionPublication . Ferreira, P. M.; Ruano, Antonio; Pestana, RuiAbstract The Portuguese power grid company wants to improve the accuracy of the electricity load demand forecast within an horizon of 48 hours, in order to identify the need of reserves to be allocated in the Iberian Market. In this work we present updated results on the identi cation of radial basis function neural network load demand predictive models. The methodology follows the principles already employed by the authors in di erent applications: the NN models are trained by the Levenberg-Marquardt algorithm using a modi ed training criterion, and the model structure (number of neurons and input terms) is evolved using a multi-objective genetic algorithm. The set of goals and objectives used in the model optimisation re ect di erent requirements in the design: obtaining good generalisation ability, good balance between one- step-ahead prediction accuracy and model complexity, and good multi-step prediction accuracy. In this work the prediction horizon was increased, the model tness assessment was altered, and the model structure search space was enlarged. Results are also presented for a predictive nearest neighbour type approach, which establishes a baseline for predictive methods comparison.
- Towards online operation of a RBF neural network model to forecast the Portuguese electricity consumptionPublication . Ferreira, P. M.; Ruano, Antonio; Pestana, RuiIn previous work the authors successfully identified a radial basis function neural network to forecast the Portuguese electricity consumption profile within a 48 hour predictive horizon. As the model is a static mapping employing external dynamics and the electricity consumption trends and dynamics are varying with time, its predictive performance degrades after a certain period. One of the simpler ways to counteract this effect is by retraining the model at certain time intervals. In this paper this methodology is investigated considering regular and irregular retraining periods. For the latter, a criterion is defined in order to trigger the retraining procedure. The results obtained are compared to a nearest-neighbour predictive approach that achieves acceptable predictive performance and operates on a sliding window of data, therefore providing some level of adaptation. Also an analysis is made in order to find the time of day where the prediction error is smaller. Globally the retraining technique provides satisfactory maintenance of predictive performance although exhibiting alternating levels.