本文为荷兰代尔夫特理工大学（作者：I. de Hoogt）的硕士论文，共100页。

电力不能有效地储存在电网中，因此需要供需平衡。输电系统运营商运营实时电力市场以获得额外的电力供应或负荷。传统上，提前一天尽可能准确地预测用电量，并且由于这些偏差涉及财务风险，因此与该电力计划的偏差应保持在最低水平。与提交的电力计划的偏差按实时电力市场结算价格结算，这对相关方是有利的。

本项目的目标是实现对电力市场价格的准确、实时、可解释的预测。为此，提出了一种具有外源输入模型结构（GFNARX）的非线性自回归模型的广义模糊神经网络公式，作为对现有模糊建模的贡献。结果表明，该模型的精度与目前流行文献中的模糊模型（Mackey-Glass混沌时间序列预测）的精度相当，同时使用了计算成本较低的方法。由于没有大量关于实时电力市场价格（REP）预测的文献能够将GFNARX预测与任何其他预测进行比较，利用时间序列文献中的另外两个模型生成比较材料：具有外源输入的季节自回归综合移动平均模型、具有外源输入的广义自回归条件异方差模型（SARIMAX-GARCH）和具有外源输入的非线性自回归模型（NARX）。评估实时电价信息是否可用于降低需求响应组合中资产的用电成本，该组合是一种基准方法，根据实时电力市场的预测状态，提出了模拟开发的冷藏库模型的冷却电机控制。根据历史数据，利用GFNARX模型预测2015年全年实时电价，与日前市场购买全部用电量的参考情况相比，累计用电量减少25.5%。将GFNARX预测结果与最新的电价观测结果进行比较，成本降低16.9%，相对成本降低10.3%。

Electricity can not be stored efficientlyin the grid, resulting in a need for demand and supply to be in balance. TheTransmission System Operator operates a real-time electricity market to acquireextra supply or load. Traditionally, electricity usage is forecast asaccurately as possible a day ahead and deviations from this electricity programare kept to a minimum as financial risk is involved with these deviations.Deviations from the submitted electricity program are settled at the real-timeelectricity market clearing price, which can be profitable to the partiesinvolved. The aim of this project is to achieve accurate, real-time andinterpretable prediction of the electricity market price. To this end theGeneralised Fuzzy Neural Network formulation of a Non-linear Auto-Regressivewith eXogenous inputs model structure (GFNARX) model structure is proposed ascontribution to existing fuzzy modelling literature. Accuracy of the proposedmodel is shown to be comparable to that of state-of-the-art fuzzy models on apopular literature benchmark, prediction of the Mackey-Glass chaotic timeseries, while using computationally cheaper means. As there is no substantialamount of literature on Real-time Electricity market Price (REP) forecastingwhich enables comparison of GFNARX predictions to any other, two other modelsfrom time series literature are used to generate comparison material: SeasonalAuto-Regressive Integrated Moving Average with eXogenous inputs and GeneralizedAuto-Regressive Conditional Heteroskedasticity (SARIMAX-GARCH) and NonlinearAuto-Regressive with eXogenous inputs (NARX). To assess whether informationabout the real-time electricity price can be used to reduce electricityconsumption costs for assets in a demand response portfolio, a benchmark methodwhich simulates control of the cooling motor of a developed cold storagewarehouse model according to the predicted state of the real-time electricitymarket, is proposed. Using the GFNARX model to predict the real-time pricethroughout the year 2015 based on historical data, a 25.5% reduction incumulative electricity costs is obtained compared to the reference case whereall electricity consumption is bought on the day ahead market. When comparingthe GFNARX prediction result to using the most recent electricity priceobservation as naïve forecast which achieves 16.9% cost reduction, a relative10.3% cost reduction is achieved.

1. 引言

2. 理论

3. 实验

4. 结果

5. 结论

更多精彩文章请关注公众号：