Search

EP-3640869-B1 - METHOD FOR PREDICTING AN ENERGY DEMAND, DATA PROCESSING SYSTEM AND RENEWABLE POWER PLANT WITH A STORAGE

EP3640869B1EP 3640869 B1EP3640869 B1EP 3640869B1EP-3640869-B1

Inventors

  • DO AMARAL BURGHI, Ana Carolina
  • HIRSCH, TOBIAS

Dates

Publication Date
20260506
Application Date
20181017

Claims (15)

  1. Method for predicting an energy demand (100) for operating a renewable power plant with a storage (101) using a machine learning device (120), wherein the method comprises: - generating a set of forecasted variables (102) based on at least one set of uncertainty-based variables (108) by using an optimization algorithm (107); and - uncertainty post-processing (105) of the set of forecasted variables (102) using machine learning techniques (104), wherein the machine learning techniques (104) are performed by a machine learning device (120) that makes use of historical data (123) relating to at least one energy demand according to a perfect forecast (148) in the past, wherein training data sets (118) for a development phase (122) within the uncertainty post-processing (105) comprise a set of deviation-to-mean-based variables (128), a set of deviation-to-persistence-based variables (130) and a set of deviation-to-perfect-based variables (136), and wherein in a development phase (122) of the uncertainty post-processing (105) the machine learning device (120) is trained by using one or more training data sets (118) as input.
  2. Method according to Claim 1, wherein the optimization algorithm (107) is performed by an optimization algorithm device (106) and a set of weather forecast variables (110) and/or a set of electricity price forecast variables (114) are used as input for the optimization algorithm device (106).
  3. Method according to any one of the preceding Claims, wherein predictive modeling techniques are used by the machine learning device (120).
  4. Method according to any one of the preceding Claims, wherein the uncertainty post-processing (105) comprises an implementation phase (124) in which the machine learning device (120) is implemented for predicting a deviation (152) of the set of forecasted (102) variables from a future energy demand according to a perfect forecast (148).
  5. Method according to any one of the preceding Claims, wherein the machine learning device (120) is trained in that deviations of at least one predicted set of variables (136) and at least one energy demand according to a perfect forecast (148) from the past are compared.
  6. Method according to any one of the preceding Claims, wherein in the development phase (122) of the uncertainty post-processing (105) a decision tree (168) is built and/or neural networks are used.
  7. Method according to any one of the preceding Claims, wherein training data sets (118) for the development phase (122) within the uncertainty post-processing (105) comprise one or more of the following sets of variables: a set of day-of-the-year-based variables (132), a set of time-step-priority-based variables (134).
  8. Method according to any one of the preceding Claims, wherein in the development phase (122) within the uncertainty post-processing (105) training data sets (118) are used by the machine learning device (120) for determining a correlation to a set of forecasted variables (102) and its deviation from an energy demand according to a perfect forecast (148).
  9. Method according to any one of the preceding Claims, wherein the development phase (122) in which a machine learning device (120) is trained is repeated, in particular with varying training data sets (118), in regular intervals.
  10. Method according to any one of the preceding Claims, wherein in an implementation phase (124) of the uncertainty post-processing (124), one or more testing data sets (135) are used as an input for a machine learning device (120), influencing decision making in the uncertainty post-processing (105).
  11. Method according to claim 10, wherein testing data sets (135) for the implementation phase (124) within the uncertainty post-processing (105) comprise one or more of the following sets of variables: a set of deviation-to-mean-based variables (128), a set of deviation-to-persistence-based variables (130), a set of day-of-the-year-based variables (132) and a set of time-step-priority-based variables (134).
  12. Method according to of any one of the preceding Claims, wherein in an implementation phase (124) within the uncertainty post-processing (105) a predicted deviation (152) of the set of forecasted variables (102) is calculated by a machine learning device (120) and wherein the predicted deviation (152) and the set of forecasted variables (102) are used as an input for an adjustment algorithm (154) performed by an adjustment algorithm device (155).
  13. Method according to any one of the preceding Claims, wherein the renewable power plant with a storage (101) is a concentrated solar power plant (180) with storage (186), a solar photovoltaic plant (200) with storage (206) or a wind power plant (220) with storage (224).
  14. Data processing system (230) adapted for carrying out the method of any one of Claims 1 to 13.
  15. Renewable power plant with a storage (101) operated according to a prediction of an energy demand (100) according to a method of any one of Claims 1 to 13 and/or comprising the data processing system (230) of Claim 14.

Description

The invention relates to a method for predicting an energy demand for operating a renewable power plant with a storage using a machine learning device. The invention further relates to a data processing system adapted for carrying out the method of the invention. The invention further relates to a renewable power plant with a storage operated according to a prediction of an energy demand according to a method of the invention and/or comprising a data processing system of the invention. A derivation of the optimization problem and its adaptation to dynamic programming is known from the scientific publication "Methodology for optimized operation strategies of solar thermal power plants with integrated heat storage" by M. Wittmann et al. The scientific publication "Optimized dispatch in a first-principles concentrating solar power production model" by M. J. Wagner et al., Applied Energy 203 (2017) 959-971, discloses the integration of a dispatch optimization model into a detailed techno-economic analysis tool, wherein the mixed-integer linear program produces an optimized operation strategy, historically determined via a heuristic. The scientific article "Variable Generation Power Forecasting as a Big Data Problem" by Sue Ellen Haupt et al. (published in IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, vol. 8, no. 2, APRIL 2017) discloses an example system that has been used for utility applications and how it has been configured to meet utility needs while addressing the Big Data issues. The presentation "Stochastic Optimization of Power Market Forecast Using Non-Parametric Regression Models" by S. Shenoy et al. (presented on the IEEE POWER & ENERGY SOCIETY GENERAL MEETING in July 2015) discloses a methodology for stochastic optimization using data-driven models. Non-parametric models of multivariate distributions based on multiple quantile regressions, built from historical data sets are used. The statistics, such as cost expectation, required for the stochastic optimization are computed numerically using these models. The "review on statistical postprocessing methods for hydrometeorological ensemble forecasting" by Wentai Li et al. (published in WIREs Water 2017, vol. 4, pages 1 to 24) discloses commonly used statistical postprocessing methods for both meteorological and hydrological forecasts. However, renewable power plants with a storage operated according to known prediction methods only service a renewable energy market, as normally accuracy of the prediction is too low in order to participate in a wholesale electricity market in which penalties for non-fulfillment of an actual energy demand are imposed and/or not being able to fulfill a contract is penalized. It is an object of the present invention to provide a reliable method for the predicting an energy demand with improved accuracy for operating a renewable power plant with a storage. This object is achieved by providing a method for predicting an energy demand for operating a renewable power plant with a storage according to claim 1. The method preferably provides a prediction of an energy demand of a specific renewable power plant, preferably based on a given energy demand of the whole electricity system. The predicted energy demand is preferably a predicted demand of production of one particular renewable power plant. Due to the uncertainty post-processing using machine learning techniques, the method of the invention can provide an automatic system for generating energy schedules. Preferably, the method provides for an unsupervised learning system. In accordance with the method, a renewable power plant with a storage is preferably operated with a high accuracy, in particular so that it can service the wholesale energy market together with conventional power plants. Alternatively, the method is suitable for predicting an energy demand of a non-renewable power system, e.g., a power-heat-power system. The energy demand according to a perfect forecast preferably is an energy demand calculated based on a perfect input, i.e., without uncertainties. The actual energy demand is influenced by various factors; inter alia, uncertainties, in particular resource-dependent uncertainties and/or market-dependent uncertainties. An example for a resource-dependent uncertainty is a weather forecast. An example for a market-dependent uncertainty is an electricity price uncertainty. In accordance with the prediction of the energy demand, preferably a time-dependent dispatch schedule is obtained. The time-dependent dispatch schedule preferably shows an amount of energy, e.g., in Megawatt MW, that is to be sent from the renewable power plant to the grid dependent on the time of the day. If less energy than given in the predicted demand is produced by a renewable energy source of the power plant at a certain time, energy from the storage of the renewable power plant is used to generate electricity which is sent to the grid. If more energy than the predicted demand is produce