An Inherently Updatable On-line Load Forecasting Technique

短期负荷预测是一项重要而复杂的任务，需要对技术和系统有深入的了解。许多基于统计和专家系统的技术产生相当准确的预测。然而，大多数统计技术在计算上是繁重的，因为这些方法需要随着条件的变化而更新。专家系统更能适应不断变化的条件，但是用明确的定量术语表达可用的专业知识（知识）并不总是那么容易，这常常导致规则不一致。本研究的目的是开发一种新的负荷预测技术，该技术结合了专家系统方法和统计技术的能力，但避免了它们的缺点。采用专家系统方法，利用现有的知识体系形成与预测点最相关的历史数据子集。建议中解释的优先向量法用于定量表达过程中涉及的任何定性因素。这种方法与其他方法有本质的不同，因为我们根据预测点的情况创建历史数据子集，而不是传统的分析整个历史数据集的方法。此外，在我们的技术中，更新功能变得自动化，因为数据子集被不断地重建，关系被重新确定以匹配每个预测点。概述了该技术最简单形式的发展，并在每周负荷预测中进行了应用演示。我们已经确定并简要讨论了需要进一步工作的领域，以使模型更加复杂以获得更好的性能，并将其扩展到执行区间预测。

Short term load forecasting is an important and complex task requiring in-depth understanding of the technique and the system. Many statistical and expert system based techniques produce fairly accurate forecast. However, most statistical techniques are computationally burdensome as these methods need updating with changing conditions. Expert systems are more adaptable to changing conditions, but it is not always easy to express the available expertise (knowledge) in clear quantitative terms, often leading to inconsistent rules. The objective of this research is to develop a new load forecasting technique which combines the capabilities of both the expert systems approach and the statistical techniques, but avoids their drawbacks. Expert systems approach is adopted to utilize the available body of knowledge to form a subset of historical data which are most pertinent to the forecast point. The priority vector method, explained in the proposal, is applied to quantitatively express any qualitative factors involved in the process. This approach is fundamentally different from the others, because we create subsets of historical data according to the conditions at the forecast point, instead of the traditional approach of analyzing the whole historical data set. Further, the updating function becomes automatic in our technique since the subset of data is continually reconstructed and the relationships redetermined to match each forecast point. The development of the technique in its simplest form is outlined and the application is demonstrated in weekly load forecast. We have identified and briefly discussed the areas which need further work to make the model sophisticated for better performance, and to extend it to perform interval forecasting.