Structuring Artificial Neural Network to Improve Energy Time Series Forecasting

Abstract

The forecasting of time series data is a classical research topic in the field of resource economics. Therefore, many forecasting studies have been being conducted in this area, including studies on predicting natural gas and crude oil demand and commodity prices. The main purpose of a forecasting study is enhancing the forecasting accuracy. A conventional way of achieving such an improvement is through the development of the forecasting model itself. However, recent studies have attempted a new approach that combines other methods. This dissertation expands the scope of such combinations. Neural network forecasting models are created in combination with data processing theory, and econometric theories. These combinations are the first approaches to improve forecasting power in neural network studies. In this study, three experiments were constructed to examine the newly proposed forecasting model. Three forecasting models were developed to reflect the characteristics of each forecasting target. An Artificial Neural Network (ANN) was used as the main forecasting model in all experiments owing to its popularly as a recent machine learning tool with a dominant forecasting capability. The first experiment included the Henry Hub weekly spot price, and the second used Korean electricity load as the forecasting target data. The Henry Hub daily spot price and daily Western Texas Intermediate (WTI) price were both used for the third experiment. In addition to the ANN, three other methods were respectively incorporated in the three experiments to improve the forecasting power of each model: wavelet decomposition, causality, and structural break. For the first experiment, the usefulness of wavelet decomposition in forecasting studies was examined. Wavelet decomposition is a pre-processing method used to filter input data, and is usually applied in the field of signal processing, such as in audio signals and images. In a time series analysis, wavelet decomposition can be used as a denoising process. Therefore, the method is considered appropriate for situations in which the overall or midterm trend of the dataset is more important than the short-term fluctuations. Thus, the Henry Hub weekly spot price was filtered using wavelet decomposition. The wavelet decomposition was adjusted appropriately for time-series forecasting, and a new forecasting model was developed. The results of this work showed about a 14% improvement in the forecasting accuracy, and verified the usefulness of wavelet decomposition in forecasting studies. The second experiment was conducted to examine the availability of the causality concept in forecasting studies. Causality is a theory typically used in the field of econometrics, and analyzes the causal relationship between two variables. Based on this method, we can identify which variable precedes another. Electricity load and temperature, which have a causal relationship, were used as variables in the experiment, and their causal relationship was used as a criterion to generate the structure of the ANN. There are currently no criteria or guidelines on how the layer structure of an ANN should be constructed. Based on the forecasting accuracy, a layer structure considering the causal relationship between variables showed the best forecasting performance. The best case showed about an 11% improvement in forecasting accuracy compared to a model not considering such causality. Therefore, we concluded that causality is useful as a criterion for structuring the layers of an ANN. The third experiment focused on forecasting both the Henry Hub daily spot price and WTI daily spot price, which have structural breaks in their dataset. A structural break is defined as a sudden change in the behavior of the data. Based on this point, a dataset can be separated as a different structure. A structural break was used as a criterion in selecting the length of the input data. This study began from a question regarding how past is the informative past in forecasting future information. Many researchers using machine-learning models believe that, the more data that are used, the better the results that are output. Thus, it is difficult to find studies handling the problem of an informative data length. The third experiment was designed to examine whether this belief is correct. The results showed that not all regimes are necessary for forecasting, and the best forecasting performance was shown for a model using a shorter dataset. An improvement in forecasting accuracy of about 7% was achieved compared to the traditional application of the longest dataset. As the results of the experiments described above, the combination of other methods such as econometric theories is effective in improving the forecasting accuracy of an ANN. In conclusion, it can be stated that, not only is the development of the forecasting model itself a way to improve the forecasting power, imposing the meaning of the variables applied to the model can also be an effective method in enhancing the forecasting power. An expansion of the practice of method combining described in this dissertation marks the first such case in research on improvements in forecasting power, and is the core contribution of this study.