| Summary: | 碩士 === 逢甲大學 === 電機工程學系 === 107 === The smart grid is considered to be a highly complex system, and its effective management is a huge challenge. Load forecasting is closely related to the development of power systems, such as energy market analysis, economic dispatch and safety assessment, and has therefore been identified as a key issue in how to effectively manage the grid. How to effectively use the vast amount of information provided by smart meters will be the focus of future research.
This paper proposes three short-term load forecasting architectures based on Long short-term memory neural networks (LSTM). There are several important improvements to improve forecasting performance. First, to improve the accuracy of neural networks, an algorithm combining wavelet transform and LSTM is proposed. Second, using the autocorrelation coefficient to filter the input data to enhance the value of the data, it can improve the efficiency of machine learning. And combined with different features to improve the overall architecture versatility. In addition, a time-series data segmentation architecture is proposed. For time-series data, data segmentation is performed under different conditions. This can effectively improve the performance of a single LSTM. The amount of data that LSTM needs to process is reduced. If it is combined with parallel computing technology, the required computing time can be greatly reduced. Finally, a time-series multi-step target prediction method is proposed, and the accuracy can still be maintained under the condition of prolonging the predicted target distance time.
This study also designed a graphical user interface (GUI) for short-term load forecasting, so that users can get the information quickly and easily.
|
|---|
