Load forecasting: a cross-field study on server and energy load forecasting Impact of temporal factors on generalization ability and performance of regression models

• Autor(a) principal: Silva, Claudio Alexandre Duarte
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: http://hdl.handle.net/10400.8/4824
• topic: Load prediction; Server load prediction; Energy load forecasting; Regression problems; Linear regression; Artificial neural networks; Support vector machines; Model generalization

The server load prediction and energy load forecasting have available a wide range of approaches and applications, with their general goal being the prediction of future load for a specific period of time on a given system. Depending on the specific goal, different methodologies can be applied. In this dissertation, the integration of additional temporal information to datasets, as a mean to create a more generalized model is studied. The main steps involve a deep literature review in order to find the most suited methodologies and/or learning methods. A novel dataset enrichment process through the integration of extra temporal information and lastly, a cross-model testing stage, where trained models for server load prediction and energy load forecast are applied to the opposite field. This last stage, tests and analyses the generalization level of the created models through the temporal information integration procedure. The created models were both oriented to short-term load forecasting problems, with the use of data from single and combined months, regarding real data from Wikipedia servers of the year 2016 in the case of server load prediction and real data regarding the consumption levels in April 2016 of the city of Leiria/Portugal for the energy load forecasting case study. The learning methods used for creating the different models were linear regression, artificial neural networks and support vector machines oriented to regression problems, more precisely the Smoreg implementation. Results prove that it is possible to tune the dataset features, e.g., granularity and time window to improve prediction results and generalization. Results from this work, as well as an optimization approach through the use of genetic algorithms, normalization effects, split ratio vs crossvalidation influence and different granularities and time windows were peer-reviewed published.