Localised high resolution forecasting for energy demand based on smart meter data

这是能源和人工智能的博士学位利用合作伙伴公用事业公司提供的智能电表数据，本博士将研究自下而上的预测模型对单个智能电表历史数据的能源需求的适用性。本博士将研究，利用合作伙伴公用事业公司提供的智能电表数据，如何使用自下而上的预测模型来预测多个尺度的能源需求。这些模型将根据来自单个智能电表的历史数据进行训练，预测将在当地/区域一级进行汇总，并与自上而下的预测准确性进行比较，以评估存在哪些效益以及效益的水平。- 空间信息允许探索场景，如Loean-Marginal定价的影响（节点定价），这是一项在美国实施并在英国进行了大量讨论的政策，基于其当前客户组合的公用事业及其客户的空间分布对盈利能力的影响。平衡本地或供应商基础内的客户组合的潜力，并确定可能的网络限制和变电站的位置，这些位置可能会随着电动汽车的使用而达到容量。设计一个最佳的交易代理对冲和购买电力给定的智能电表信息的投资组合是一个可能的最终应用的研究，将汇集一些个别股的project.Methods：作为一个起点，国家的最先进的集成方法将作为一个基准进行测试，然后探索的潜力，深度学习方法，特别是长短期记忆神经网络，以改善结果。强化学习和基于代理的模型也将在整个项目过程中进行探索。

This is a PhD in Energy and Artificial Intelligence. Using smart meter data provided by the partner utility company, this PhD will investigate the applicability of bottom-up forecasting models for energy demand trained on historical data from individual smart meters.This PhD will investigate, using smart meter data provided by the partner utility, how bottom-up forecasting models can be used to predict energy demand at a number of scales. These models will be trained on historical data from individual smart meters, and the forecasts will be aggregated at a local/regional level and compared with the accuracy of top down forecasts to assess what benefits exist and what level they occur.Particular avenues of exploration could be:- Spatial information allowing exploration of scenarios such as the impact of Locational Marginal Pricing (nodal pricing), a policy implemented in the United States and much discussed in the UK, on a utility based on its current portfolio of customers and the impact of the spatial distribution of its customers on profitability.- the potential for balancing portfolios of customers, locally or within the suppliers base, and identifying locations of possible network constraints and substations likely to reach capacity with the uptake of electric vehicles.- Design of an optimal trading agent for the hedging and purchasing of electricity given the portfolio of smart meter information is a likely final application of the research that would bring together a number of the individual strands of the project.Methods:As a starting point, state-of-the-art ensembling methods will be tested as a benchmark before exploration of the potential for deep learning methods in particular Long-Short-Term-Memory Neural Networks to improve results. Reinforcement learning and agent-based models will also be explored throughout the course of the project.