Uma metodologia para tratamento de dados de curvas de carga baseada em técnicas de inteligência artificial

Abstract

Data quality is critical in the short-term load forecasting. Frequently, load data show aberrant values (outliers), discontinuities, and gaps (missing data) caused by the abnormal operation of the electrical system or failures and problems in the measurement system. The presence of corrupted data impairs specification of load forecasting models and consequently affects the quality of predictions obtained. Therefore, the construction of a load prediction model must be preceded by a data processing step. This dissertation presents a methodology based on statistical methods and artificial intelligence for the treatment of load data. Throughout the dissertation are presented the methods used and how each of them is employed in the identification and correction of the main types of errors frequently found in the load data. In addition, computational experiments were conducted with load data from the National Interconnected System in order to evaluate the ability of the proposed methodology to clean and recover the original patterns of corrupted load curves. In the experiments performed the load curves were artificially corrupted by means of statistical simulation and later treated by the proposed methodology. The results show the good adherence of the load curves resulting from the data cleaning process to their original uncorrupted profiles. Computational experiments were conducted with real data from the National Interconnected System (SIN) to evaluate the ability of the proposed methodology to clean load data and recover the original patterns of corrupted load curves. In the experiments, the load curves were artificially corrupted and then filtered by the proposed methodology. The results show the good adherence of the load curves resulting from the data cleaning process to their original uncorrupted profiles.