Systems and methods for using machine learning to predict critical constraints

Abstract

A computer implementation method and a computer program product are provided for predicting the required committed capacity of an electric utility. The method includes the steps of: (a) performing probabilistic optimization of raw data to generate the total committed capacity from conventional thermal units as target data, wherein the raw data includes power grid operating conditions; (b) combining the total committed capacity from conventional thermal units with raw features and designed features to generate training data; (c) training a machine learning model for predicting the required committed capacity of an electric utility using the generated training data; (d) predicting the required committed capacity of a power company using the trained machine learning model; and (e) running an extended version of a deterministic output allocation optimization model based on the predicted, required committed capacity of an electric utility. The computer program performs the steps described above.

Description

Cross - reference to related applications

[0001] This application claims priority to U.S. Patent Application No. 63 / 487,942, filed on March 2, 2023, entitled "Systems and Methods for Using Machine - Learning to Predict Critical Constraints", which is hereby incorporated by reference in its entirety.

Technical Field

[0002] The present disclosure relates to power generation in an electric utility, and more particularly, to predicting the firm capacity required by an electric utility.

Background Art

[0003] A vertically integrated utility ("VIU") creates a day - ahead ("DA") unit commitment and power generation dispatch schedule based on predictions of future load demand, resource availability, fuel prices, etc. These schedules include the on / off status and output of generators, and the charge / discharge patterns of energy storage assets. Since the commitment (on / off) status cannot be adjusted in real - time in response to an emergency notice, by scheduling in advance, a slow - start generator (e.g., a steam turbine or a combined - cycle plant) can have sufficient advance notice and as a result, can follow the planned operation. However, the optimality of these day - ahead commitment and output dispatch schedules depends on how well the uncertainty of the future power grid state is considered in real - time with respect to what actually occurs later.

[0004] In particular, day-ahead forecasts for variable renewable energy sources ("VRE"), such as solar and wind power, are often inaccurate, and their real-time generation can deviate significantly from the forecast. These VRE deviations also cause fluctuations in the net load (demand - VRE generation) that other non-VRE resources must meet. The uncertainty of net load occurring in real time makes it difficult to schedule an appropriate amount of non-VRE generation in advance, especially for start-delay generators, where real-time commitment (on / off) status cannot be corrected. One way to address this problem is to overschedule non-VRE generator capacity beyond sufficient on the day before to account for the range of possible VRE generation. However, this approach is very costly and inefficient.

[0005] Instead, current practice involves power grid operators using fast-start generators (e.g., combustion turbine plants) or reducing variable real-time energy (VRE) to adjust for the mismatch between real-time demand and generation. This method of compensating for resource shortages increases the cost of the equilibrium operation, which is the cost difference between the previous day's power allocation and the real-time power allocation. However, over-reliance on fast-start generators (which are expensive) and VRE reductions is also very costly and wasteful. Therefore, to meet net load demand in real time, there are costs and risks associated with both over-scheduling and under-scheduling generator capacity the previous day.

[0006] As more VRE resources are installed and operational, net load fluctuations will only increase, making grid operation more complex. Optimal output allocation schedules can result in higher energy costs and reduced grid reliability. To address the uncertainty of net demand, scheduling algorithms are still needed that can balance slow-start generation scheduled the day before with fast-start generation available in real time. Furthermore, there remains a need to modify current grid operations to schedule appropriate ramp-feasible generator capacity the day before to cost-effectively respond to real-time net demand fluctuations. [Overview of the project]

[0007] In the first example, a computer implementation method 500 for predicting the required committed capacity of an electric utility 100 includes performing a probabilistic optimization 402 of raw data 304 to generate the total committed capacity from conventional thermal units as target data. The raw data 304 includes power grid operating conditions 306. The method 500 further includes combining the total committed capacity from conventional thermal units with raw features 304 and designed features to generate training data. The method 500 further includes training a machine learning model 404 for predicting the required committed capacity of the electric utility 100 using the generated training data. The method 500 further includes predicting the required committed capacity of the electric utility 100 using the trained machine learning model 404. The method 500 further includes running an extended version of a deterministic output allocation optimization model 406 based on the predicted, required committed capacity of the electric utility 100.

[0008] In the second example, a non-temporary computer-readable medium 204 includes a control program 302 for the required committed capacity (of the electric utility). By one or more processors 202 executing the control program 302, one or more computing devices 116 are configured to perform a probabilistic optimization 402 of the raw data 304 to generate the total committed capacity from conventional thermal units as target data. The raw data 304 includes power grid operating conditions 306. By one or more processors 202 executing the control program 302, one or more computing devices 116 are configured to generate training data by combining the total committed capacity from conventional thermal units with the raw features 304 and designed features. Once one or more processors 202 have executed the control program 302, one or more computing devices 116 are configured to use the generated training data to train a machine learning model 404 to predict the required committed capacity of the electric utility 100. When one or more processors 202 execute control programming 302, one or more computing devices 116 are configured to predict the required committed capacity of the electric utility 100 using a trained machine learning model 404 for predicting the required committed capacity of the electric utility 100. When one or more processors 202 execute control programming 302, one or more computing devices 116 are configured to execute an extended version of the deterministic output allocation optimization model 406 based on the predicted, required committed capacity of the electric utility 100.

[0009] In the third example, the computing device 116 includes memory 204, a processor 202 coupled to memory 204, and programming 302 within memory 204. By the processor 202 executing programming 302, the computing device 116 is configured to perform probabilistic optimization 402 of raw data 304 to generate total committed capacity from conventional thermal units as target data. The raw data 304 includes power grid operating conditions 306. By the processor 202 executing programming 302, the computing device 116 is configured to generate training data by combining the total committed capacity from conventional thermal units with raw features 304 and designed features. Once the processor 202 has executed programming 302, the computing device 116 is configured to use the generated training data to train a machine learning model 404 to predict the required committed capacity of the electric utility 100. When processor 202 executes programming 302, computing device 116 is configured to use a trained machine learning model 404 to predict the required committed capacity of the electric utility 100. When processor 202 executes programming 302, computing device 116 is configured to run an extended version of the deterministic output allocation optimization model 406 based on the predicted, required committed capacity of the electric utility 100.

[0010] The additional purposes, advantages, and novel features of the example are described in part in the following description and in part may become apparent to those skilled in the art by examining the following and the accompanying drawings, or by generating or manipulating the example. The purposes and advantages of this disclosure are realized and achieved by the methods, means, and combinations specifically indicated in the accompanying claims.

[0011] The figures in the drawings show one or more embodiments as examples only, not as limitations. In the figures, similar reference numerals refer to identical or similar elements. [Brief explanation of the drawing]

[0012] [Figure 1] This disclosure illustrates an exemplary system for managing an energy system according to embodiments of this disclosure. [Figure 2] Figures 4 and 5 show schematic diagrams of exemplary computing devices according to embodiments of the present disclosure that can implement the methods shown in these figures. [Figure 3] Figure 1 is a high-level functional block diagram of a system for managing an energy system, showing the components of the controller, energy system, and load and price forecasting system, in order to predict the required committed capacity of an electric utility. [Figure 4] This is a flowchart illustrating an extended deterministic optimization according to one embodiment. [Figure 5] This flowchart shows an overall method for predicting the required committed capacity of electricity utilities. [Figure 6A] The plots show the capacity constraint performance against the minimum conventional capacity constraint, the pre-planned conventional capacity ("DA") calculated using a probabilistic method, and the DA-planned conventional capacity calculated using an extended deterministic model. [Figure 6B]This paper demonstrates optimized real-time power generation using current practices (e.g., minimal conventional capacity constraints), probabilistic methods, and extended deterministic models. Parts List 100 Electricity Utilities 101 Energy Management Systems 102 Energy Systems 104 Renewable Power Generation Systems / Sources 106 Non-Renewable Power Generation Systems / Sources 108 Energy Storage Systems 110 Power Applications 112 Controllers 114 Load and Price Forecasting Systems (Multiple Possible) 116 User Computing Devices 118 Networks 202 Processors 204, 322 Memory 206 Network Communication Interfaces 208 Input Devices 210 Output Devices 302 Required Capacity Control Programming 304 Raw Data 306 Power Grid Operating Conditions 308 Solar or Wind Power 310 Weather Forecast 312 Net Load Demand per Hour 314 Net Replacement Schedule 316 Region 318 Net Storage Schedule 320 Total Capacity per Hour of Start-Up Delay Units 324 Historical Storage Conditions 326 Predictive Storage Control Programming 328 Total Capacity per Hour of Conventional Thermal Units 330 Machine learning models ("ML") for predicting optimal storage behavior 400 Extended deterministic optimization methods 402 Probabilistic optimization 404 Machine learning models for predicting capacity 406 Extended deterministic models 500 Method 602 Minimum conventional capacity constraint 604 Day before ("DA") planned capacity using probabilistic methods 606 Day before planned capacity using deterministic models [Modes for carrying out the invention]

[0013] In the embodiments for carrying out the following inventions, numerous specific details are given in order to provide a complete understanding of the relevant teachings. However, it should be obvious to those skilled in the art that embodiments can be practiced without such specific details. In other examples, well-known methods, procedures, components, and / or circuits have been described at a relatively high level without detail in order to avoid unnecessarily obscuring the embodiments of these teachings.

[0014] Unless otherwise indicated, any embodiment can be combined with any other embodiment. In particular, Figures 1-6 and the related text are all combinable with each other.

[0015] As used herein, the term “coupled” means any logical, physical, electrical, or optical connection, link, etc., through which electricity, output, signal, or light generated or supplied by one system element is transferred to another coupled element. Unless otherwise stated, coupled elements or devices are not necessarily directly connected to one another and may be separated by intermediate components, elements, or communication media that can modify, manipulate, or carry electricity, output, signal, or light.

[0016] This specification describes how machine learning can be used to predict critical constraints on utility power generation facilities to efficiently meet customer loads and integrate the predicted critical constraints into a mathematical optimization model.

[0017] The methods described herein can help utilities schedule power generation facilities to meet their loads while minimizing operating costs. Traditionally, this would be achieved using mathematical optimization algorithms that are inherently deterministic. Probabilistic programming techniques yield better results than deterministic algorithms, but at the cost of additional computational complexity and time. While the market is receptive to the more optimal results produced by probabilistic methods, the computation time is not ideal for some use cases.

[0018] The methods described herein result in outcomes at a level of speed and computational complexity equivalent to deterministically derived solutions. The methods conventionally use deterministic optimization for economic output distribution optimization of unit commitment (e.g., how a utility plans the schedule of its units to meet the demand of the power grid). In particular, deterministic approaches use a single set of future predictions of load (e.g., forecasts), future predictions of price, and future predictions of the availability of renewable facilities. Deterministic approaches need to have some future prediction in order to schedule power generation facilities in advance.

[0019] Probabilistic optimization is based on the same concept but uses multiple different scenarios of future predictions. In particular, probabilistic optimization optimizes not just a single set of predictions but for a series of load predictions. This approach takes into account, for example, that the load may vary over the next 24 hours by using past errors or the best available information.

[0020] Probabilistic methods are computationally difficult and take more time but result in better outcomes than deterministic methods. Probabilistic methods can generate a range of predictions, and if the future lies within that range, the solution is superior to deterministic methods that consider only a single prediction and whose results deviate significantly from that prediction. However, since decisions about the power grid need to be made the day before, the time available to solve complex models of mathematical equations is limited. As the computation time increases, probabilistic methods become unsuitable for predicting critical utility power generation facilities. For these reasons, the electric utility industry currently mainly uses deterministic methods, if not exclusively.

[0021] The methods described herein combine the advantages of probabilistic and deterministic methods. In particular, the methods described herein are based on probabilistic optimization and use machine learning to learn the generation capacity schedule, the units to be scheduled, and how much capacity to obtain from each unit, taking into account predicted future forecasts. The method uses a generation capacity schedule predicted by a machine learning model that replicates the behavior of probabilistic optimization as input for a deterministic optimization model. In other words, the methods described herein use machine learning to bridge the gap between probabilistic and deterministic optimization.

[0022] More specifically, the methods described herein train a machine learning model with carefully selected important constraints that affect the determination of the reserve capacity generated by a more complex probabilistic model.

[0023] The machine learning model then needs to be started one day before actual operation to predict the required approximate capacity from conventional thermal power generation facilities such as nuclear power plants, coal plants, and natural gas plants. The approximate capacity in the context of this description means that the capacity from conventional thermal power generation facilities is scheduled to turn on and generate electricity at a specific time at least one day before the actual operation of the conventional thermal power generation facilities so that the conventional thermal power generation facilities can start and operate with sufficient time. The approximate capacity in the context of this description is the approximate capacity of a start-up delay type thermal power generator. The model uses power grid operating conditions (e.g., net demand per hour, net exchange, etc.) to form its prediction and generates a total heat capacity value at the desired time interval. This prediction is sent to a deterministic optimization algorithm as a capacity constraint, which helps to produce high-quality results comparable to probabilistic methods but with a speed similar to that of deterministic methods.

[0024] Probabilistic optimization can be run once a month or once a week to obtain updated data for the machine learning model. The machine learning model then transfers its learning to a deterministic model that runs more frequently.

[0025] Here, we will refer in detail to the example shown in the attached diagram and described below.

[0026] Figure 1 shows an exemplary system 101 for managing an energy system 102. The energy system 102 may include one or more renewable power generation systems 104, one or more non-renewable power generation systems 106, and / or one or more energy storage systems 108 that can deliver power to power applications 110 (e.g., power consumers or power grids). The system 101 may also include one or more controllers 112, one or more load and price forecasting systems 114, and / or one or more user devices 116 connected to each other via a network 118.

[0027] Network 118 may include one or more of various types of networks for the transmission of information, such as cellular networks (e.g., 2G, 3G, 4G, or 5G), satellite networks, Wi-Fi networks, WiMAX networks, Bluetooth networks, near-field communication (NFC) networks, low-power wide-area networks (LPWAN) networks, mobile networks, terrestrial microwave networks, wireless ad-hoc networks, Ethernet networks, telephone networks, power line communication (PLC) networks, coaxial cable networks, and / or fiber optic networks. Network 118 may include wired or wireless networks. Network 118 may include personal area networks, local area networks, metropolitan area networks, wide area networks, global area networks, space networks, or any other type of computer network that can use data connectivity between network nodes. In some examples, network 118 may include Internet Protocol (IP) based networks.

[0028] The renewable power generation system 104 may include renewable energy sources such as solar and wind power, which may be intermittent and less reliable compared to fossil fuels.

[0029] The non-renewable power generation system 106 may include renewable energy sources such as oil, coal, natural gas, hydroelectric power systems, or other types of non-renewable power generation systems.

[0030] The energy storage system 108 may include batteries or other devices capable of storing and releasing energy. To improve resilience, the energy storage system 108 can store energy from the energy system 102 when the output from power sources (e.g., 104, 106) is high. Later, the energy storage system 108 can output and distribute energy to the power application 110 when demand is high or when the output from power sources (e.g., 104, 106) is not keeping up with demand. Furthermore, events may occur when the connected load or operating demand load of the power application 110 is excessive, or when the power grid is unstable, for example, during extreme weather. By storing energy from power sources (e.g., 104, 106) and then outputting and distributing that energy during such events, the energy storage system 108 can continue to output and distribute the required flow of power to the power application 110.

[0031] Power application 110 may include distribution networks such as transmission grids or smaller local loads such as backup power systems for facilities such as hospitals, manufacturing sites, homes, or other suitable establishments. Power application 110 may output and distribute AC or DC power for on-grid or off-grid applications, including commercial, industrial, or residential applications. Power application 110 may output and distribute power to buildings, electric vehicle charging stations, etc., which include various electrical loads that consume AC or DC power. Power application 110 may be a front-of-the-meter system owned or operated by a utility company, or a behind-the-meter system that directly supplies electricity to buildings and homes.

[0032] The user device 116 may include any type of computing device configured to perform one or more of the embodiments described herein (for example, to optimize energy output distribution and / or for anomaly detection in an energy storage system). For example, the user device 116 may include at least one processor 202 and a memory 204 (Figure 2) that, when executed by the at least one processor 202, stores instructions causing the at least one processor 202 to perform one or more of the embodiments described herein. The user device 116 may include, for example, a computer, laptop computer, desktop computer, mainframe computer, tablet, smartphone, mobile phone, mobile device, server device, client device, automotive electronics device, augmented reality headset, smartwatch, Internet of Things (IoT) device, or any other type of computing device. In some examples, the user device 116 may be configured to receive data from various sources (for example, via network 118) and / or manage the energy system 102.

[0033] Figure 2 shows an exemplary computing device 116 consistent with several embodiments of the present disclosure that can implement methods 400, 500 of Figures 4-5. The computing device 116 may include, for example, at least one processor 202, at least one memory 204, at least one network communication interface 206, one or more input devices 208, and / or one or more output devices 210. The devices described herein (e.g., user device 116), controller 112, and / or other computing devices may similarly include these components and / or be similarly implemented. In some examples, the computing device 116 of Figure 2, which includes one or more of the above-described components, may be implemented using virtualization and / or cloud computing technologies.

[0034] The processor 202 may execute instructions of a computer program to perform any of the functions described herein. The processor 202 may include, for example, an integrated circuit, a microchip, a microcontroller, a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a field-programmable gate array (FPGA), or other units suitable for executing instructions or performing logical operations. The processor 202 may include a single-core processor or a multi-core processor (e.g., a dual-core, quad-core, or any desired number of cores).

[0035] The memory 204 may include a non-temporary computer-readable medium that, when executed by at least one processor (e.g., processor 202), can store instructions that cause at least one processor 202 to execute one or more processes as described herein.

[0036] The network communication interface 206 may include, for example, a network card, a modem, etc., and may be configured to provide data communication (e.g., bidirectional data communication) with a network (e.g., a network). The network communication interface 206 may be a wireless communication interface, a wired communication interface, or a combination of the two.

[0037] The input device 208 may include, for example, a keyboard, mouse, touchpad, touchscreen, one or more buttons, joystick, microphone, and / or any other device configured to detect and / or receive input. In some examples, the input device 208 may include one or more of various types of sensors, such as an image sensor, temperature sensor, humidity sensor, position sensor, or any other type of sensor.

[0038] The output device 210 may include, for example, a light indicator, a light source, a display (e.g., a light-emitting diode (LED) display, an organic light-emitting diode (OLED) display, a liquid crystal display (LCD), or a dot matrix display), a screen, a touchscreen, a speaker, headphones, a device configured to provide haptic cues, a vibrator, and / or any other device configured to provide output.

[0039] Memory 204, when executed by at least one processor (e.g., processor 202), may store instructions that cause at least one processor 202 to execute one or more processes as described herein. Instructions may include, for example, software instructions, computer programs, computer code, executable instructions, source code, machine instructions, machine language programs, or any other type of instructions for computing devices. Instructions may be based on one or more desired programming languages ​​of various types and may include (e.g., embody) various processes for optimizing energy output distribution and / or detecting anomalies in an energy storage system, as described herein.

[0040] In the context of this description, the functions of the computer implementation method 500 for predicting the required committed capacity of an electric utility may be performed by the processor, controller 112, and / or user device 116 of the load and price forecasting system 114. The program instructions for the method for predicting the required committed capacity of an electric utility may be stored in the memory of the load and price forecasting system 114, controller 112, and / or user device 116.

[0041] Figure 3 shows a high-level functional block diagram of the energy management system 101 of Figure 1, showing the components of the controller 112, energy system 102, and load and price forecasting system 114, which are connected to each other via a network 118 to predict the required committed capacity of an electric utility. As shown, the energy system 102 may include one or more renewable power sources 104, one or more non-renewable power sources 106, and / or one or more energy storage systems 108, or a combination thereof, which can deliver power to power applications 110 (e.g., power consumers or the power grid).

[0042] The controller 112, energy storage nodes 105A-105N, energy system 102, load and price forecasting system 114, power application 103, and other components of system 101 can communicate via network 118. Network 118 may be a local area network, a wide area network, or a combination thereof. For example, controller 112 can be connected via network 118 to load and price forecasting system 114, power sources 104 and 106, energy storage system 108, and power application 110.

[0043] The controller 112 includes a network communication interface 206 configured for wired or wireless communication over a network 118. The controller 112 further includes a memory 204 and a processor 202 coupled to the network communication interface 206 and the memory 204. As shown in the figure, the memory 204 of the controller 112 is configured to store essential capacity control programming 302 and raw data 304 such as power grid operating conditions 306, net load demand per hour 312, and net replacement schedule 314.

[0044] The controller 112 is configured to receive energy from the energy system 102 and store additional raw data 304 in memory 204 or in a separate memory 322 of the energy system 102, such as solar or wind power generation 308, weather forecast 310, region 316, net storage schedule 318, total hourly capacity 320 of load-following startup delay units, and observed past storage conditions 324. The net storage schedule 318 may include an optimized schedule and a predicted schedule.

[0045] As further shown in Figure 3, the memory 322 of the energy system 102 (or the memory 204 of the controller 112) can be configured to store the predictive storage control programming 326 and the total hourly capacity 328 of the conventional thermal units scheduled for the previous day ("DA") cycle.

[0046] The memory 322 of the energy system 102 (or the memory 204 of the controller 112) may be further configured to store a machine learning model ("ML") 330 for predicting optimal storage behavior. The machine learning model 330 for predicting optimal storage behavior may be a separate module or part of the predictive storage control programming 326.

[0047] Figure 4 is a flowchart illustrating an extended deterministic optimization method 400 applied to a computer implementation method for predicting the required committed capacity of an electric utility, according to one embodiment. The extended deterministic optimization method 400 shown in Figure 4 combines the advantages of probabilistic and deterministic methods by performing probabilistic optimization using multiple scenarios of raw data 304 (e.g., future predictions of power grid operating conditions 306) to generate total committed capacity from conventional thermal units, and combines the total committed capacity from conventional thermal units with raw features 304 and designed features to generate training data, and uses the generated training data to train a machine learning model to predict the required committed capacity of an electric utility. Based on the predicted, required committed capacity of the electric utility, the extended deterministic optimization method 400 uses a probabilistic optimized power generation capacity schedule as input for an extended version of a deterministic power allocation optimization model.

[0048] The extended deterministic optimization method 400 shown in Figure 4 may include four separate models: a probabilistic optimization 402, a machine learning model 330 (shown in Figure 3) for predicting optimal storage behavior, a machine learning model 404 for predicting the required committed capacity of the electricity utility, and an extended deterministic model 406. Training the machine learning model 404 for predicting the required committed capacity of the electricity utility does not require storage prediction.

[0049] Figure 5 is a flowchart of the overall method 500 for predicting the required committed capacity of an electric utility. Method 400 may be performed by the controller 112. Alternatively, method 500 may be performed by the processor of the load and price forecasting system 114 and / or user device 116.

[0050] In step 510 of Method 500, the controller 112 performs probabilistic optimization of the raw data 304 (e.g., initial input features) to generate the total committed capacity from conventional thermal units as target data. Conventional thermal units may include, for example, nuclear power plants, coal plants, or natural gas plants, combined cycle ("CC") sources, combustion turbines ("CT"), or steam turbines ("ST").

[0051] The raw data 304 may include power grid operating conditions 306 (Figure 3).

[0052] In a particular embodiment, as shown in Figure 3, the raw data 304 may further include solar or wind power generation 308, weather forecasts 310, multiple sets of forecasts of net load demand per hour 312, a net exchange schedule 314, a region 316, and a net storage schedule 318.

[0053] The net storage schedule 318 may include an optimized schedule and a predicted schedule.

[0054] In certain embodiments, as shown in Figure 3, the raw data 304 may further include the total capacity 320 of the time-based load-following startup delay units, which are planned by probabilistic optimization.

[0055] Initial input features can include hourly statistics of net load predictions. Net load scenarios can be calculated from load and solar energy predictions used in a probabilistic model by calculating the mean, minimum, maximum, range, variance, skewness, and kurtosis of load and solar energy predictions hourly.

[0056] The initial input features can include the average total load per hour (MWh), the net exchange schedule per hour (= export-import in MWh units) used in the stochastic optimization, and the DA net storage schedule per hour resulting from the stochastic optimization. Net storage schedule = discharge-charge in MWh units, averaged across all scenarios for each pump storage and battery unit, and summed up for all units for the total per hour.

[0057] Since fluctuations in net demand from one year to the next are recognized, Method 500 is not limited to training data equivalent to one year. To increase the amount of training data, Method 500 can artificially generate four additional sets of load scenarios by randomly increasing or decreasing the hourly load within a given range, thereby randomly modifying the net load. The newly designed features lead to better predictions of conventional capacity. The artificial data generation can also be used to model the growth of demand and renewable generation over the next few years. Alternatively, Method 500 can use historical data instead of artificially generating additional data.

[0058] In step 512 of method 500, the controller 112 generates training data by combining the total guaranteed capacity from the conventional thermal unit with the raw feature 304 and the designed feature.

[0059] In step 514 of Method 500, the controller 112 uses the generated training data to train a machine learning model to predict the required committed capacity of the electric utility.

[0060] Training machine learning models can include neural networks (PyTorch), random forests (scikit-learn), and gradient boosting (LightGBM and XGBoost).

[0061] Training a machine learning model to predict the required committed capacity for electricity utilities may include a step of designing additional input features from the raw data 304.

[0062] From the initial input set (e.g., raw data 304), the controller 112 can automatically generate a large list of more than 300 designed features. Examples of more than 300 designed features include raising a variable to a different power (e.g., x 2 , x 3 This may include, for example, different combinations of ratios (e.g., x² / x³, x² / x⁴, etc.), different delay values ​​(e.g., x at t-5 to predict y at t, x at t-4 to predict y at t, etc.), and different variable transformations (e.g., log(x)).

[0063] Training a machine learning model to predict the required committed capacity for electricity utilities may include steps to select additional input features and a subset of features from the raw data 304.

[0064] The more than 300 designed features can be reduced to about 80 features by removing low-information, high-null, and single-value features, as well as highly correlated features (for example, the FeatureTools package includes functions for all four of these reductions). From the remaining approximately 80 features, Method 400 selects the top 10 features that have the best mutual information score with the target data for training (e.g., computed by scikit-learn).

[0065] The target data may include the total capacity of conventional thermal units per hour (e.g., guaranteed capacity per hour) as planned by probabilistic optimization in the previous day ("DA") cycle. The guaranteed capacity per hour is the sum of the DA capacities, independent of the scenarios of nuclear, coal, CC, CT, and ST units. This calculation excludes hydroelectric units, solar energy units, pump-storage hydroelectric units, and battery storage units.

[0066] Mutual information captures the nonlinear dependence between input features and the target variable.

[0067] A subset of selected features (e.g., final input features, all per hour) may include mean net demand, delayed mean net demand (e.g., t-1 to t-5), mean net demand / mean total demand, month within the year (e.g., an integer), Is_weekday (e.g., a binary indicating whether the day is a weekday as opposed to a weekend), net storage schedule (=discharge-charge), and net exchange schedule (=export-import).

[0068] The step of selecting a subset of features may include measuring the dependency between each feature and a target variable representing the total conventional capacity expected by probabilistic optimization, ensuring that the dependency is equal to zero when two random variables are independent, capturing nonlinear dependencies, removing low-information features, reducing highly correlated features, performing Lasso regression, Ridge regression, and selecting top features that have the best mutual information scores with the target data.

[0069] Probabilistic optimization (step 510 in Figure 5) is sometimes used to generate training data for machine learning models to predict the required committed capacity of electric utilities. Probabilistic optimization is used only when new training data is needed for generation.

[0070] In step 516 of method 500, the controller 112 predicts the required committed capacity of the electric utility using a trained machine learning model for predicting the required committed capacity of the electric utility.

[0071] Machine learning models for predicting the required committed capacity of electricity utilities are used daily to generate power generation capacity constraints for conventional heat sources in extended deterministic power allocation optimization models.

[0072] In step 518 of Method 500, the controller 112 runs an extended version of the deterministic output allocation optimization model based on the predicted and required committed capacity of the electric utility. Specifically, the controller 112 uses the generated, predicted, and required committed capacity schedule from the stochastic optimization (steps 510-516) as input for the extended version of the deterministic output allocation optimization model.

[0073] Generally, a deterministic model may be a mixed-integer linear optimization model ("MILP") of a unit commitment / economic output allocation problem that seeks to minimize the cost of committing to and allocating output to power generation units, given a set of constraints on resources (e.g., power generation systems, storage, demand response, resource mixing, transmission systems, etc.) in order to ensure reliability and economic efficiency. All future predictions can be interpreted as certainties, and the deterministic model optimizes power generation against a set of certain future predictions such as load, energy availability, and price. These predictions may be single-point estimates over time that are assumed to occur with 100% probability.

[0074] The extended deterministic model is similar to the deterministic model but adds a constraint on power generation capacity for conventional heat sources, including, but not limited to, combined cycle ("CC") sources, combustion turbines ("CT"), and / or steam turbines ("ST"). The addition of this power generation capacity constraint allows the deterministic model to intelligently mimic the behavior of probabilistic optimization. This constraint can be added as a way to bring back uncertain information (e.g., net load uncertainty) that probabilistic optimization involves, without needing to use uncertain future forecasts (i.e., scenarios), as in the probabilistic model, based on predicted total conventional capacity rather than CC capacity, because conventional capacity as a whole directly corresponds to the net load.

[0075] The predicted and required committed capacity of the electric utility is used to extend conventional deterministic power allocation optimization models by running them using capacity forecasts as the minimum constraint on conventional capacity in the previous day ("DA") cycle.

[0076] The following equations (1a) to (1j) represent the formulation of a deterministic unit commitment ("UC"). The extended deterministic model minimizes the total cost of meeting net electricity demand (demand - planned VRE production) while adhering to generator operating limits and market clearing constraints.

[0077]

number

[0078]

number

[0079]

number

[0080] The “extended” component of the deterministic model is captured by equation (1j), which enforces a minimum total capacity constraint on conventional thermal units. Equation (1j) ensures that, given uncertainty in the predicted net demand, the total capacity planned from conventional thermal units on the previous day ("DA") reaches what the stochastic DA-UC would plan from these units (or a prediction of what the stochastic DA-UC would plan). By including the minimum total capacity constraint, it is ensured that a sufficiently flexible capacity is planned in the operating generators to avoid uncertainty in VRE generation and net demand. In other words, the planned operating capacity provides not only the variable reserve needed to manage the actual net demand fluctuations from the previous day's forecast in the most cost-effective way, but also sufficient capacity to supply the expected consumption. This constraint also allows the extended deterministic model to violate the minimum capacity requirement when it is reasonably cost-effective to do so. For example, slack variables, dev t This represents a deviation from the minimum capacity requirement, and the penalty cost evaluates the associated cost per MW of not meeting the requirement. Total cost of violating the constraint (dev t The penalty (x) is included in the objective function. Violations are permitted to ensure that the model satisfies the requirements in the most cost-effective way. For example, if the planned capacity is 1 MW short of the minimum threshold, then planning an additional 1 MW (which would require starting another generator) may not be practical. Violations give the model the flexibility to avoid additional startup costs simply to meet the capacity threshold specified by the constraint.

[0081] The added power generation capacity constraint may be a capacity constraint notified by machine learning ("ML"). In particular, Method 500 may include training a machine learning model to predict the storage schedule by using observed past storage states. The storage schedule is a critical input for capacity prediction, and the capacity prediction model is trained on actual storage schedules from stochastic optimization. However, to actually predict capacity, the storage schedule input cannot be obtained from stochastic optimization due to computational constraints on the stochastic optimization model. Therefore, Method 500 uses a second machine learning model to predict the charging and discharging behavior of storage.

[0082] The predicted storage is used as an input feature for a machine learning model to predict the required committed capacity of the electricity utility. The machine learning model for predicting optimal storage behavior generates a storage schedule for test data by training the machine learning model for predicting the required committed capacity of the electricity utility to predict the charging and discharging behavior of the storage.

[0083] A machine learning model for predicting optimal storage behavior is trained daily to generate predictions of storage behavior. The predicted storage behavior is used as input features for a machine learning model to predict the required committed capacity of the electricity utility.

[0084] An extended version of the deterministic output allocation optimization model is run daily to generate unit commitments and output allocation plans for customers by using the predicted and required committed capacity of the electric utility as input. Specifically, the extended version of the deterministic optimization model outputs an optimized storage schedule, while the stochastic optimization uses the optimized storage schedule as an input feature.

[0085] Machine learning models used to predict the required committed capacity for electricity utilities can be improved by retraining them with updated data.

[0086] A machine learning model for predicting the required committed capacity of electricity utilities can predict total heat capacity values ​​at predetermined time intervals. Total heat capacity values ​​may include, for example, the capacity of thermal power generation facilities such as nuclear power plants, coal plants, or natural gas plants, combined cycle ("CC") sources, combustion turbines ("CT"), or steam turbines ("ST").

[0087] Machine learning models for predicting the required committed capacity of electricity utilities use grid operating conditions to predict total heat capacity values ​​at predetermined time intervals. These grid operating conditions may include, for example, net demand per hour and net exchange.

[0088] Figure 6A shows plots of capacity constraint performance for the minimum conventional capacity constraint 602, the previous day's ("DA") planned conventional capacity 604 using a probabilistic method, and the DA planned conventional capacity 606 using an extended deterministic model.

[0089] Figure 6B shows optimized real-time power generation using current practices (e.g., minimal conventional capacity constraints), probabilistic methods, and an extended deterministic model. Adding capacity constraints to the deterministic model achieves the desired effect; that is, the extended deterministic model yields lower-cost unit commitment and output allocation solutions than the original deterministic model, in part the time required for the probabilistic model to solve. The extended deterministic model reduces costs by -0.7% compared to the conventional deterministic model for current power generation mix scenarios. This trend also holds for future power generation mix scenarios with increased solar energy and storage.

[0090] Probabilistic models, like deterministic models, may also be mixed-integer linear optimization models ("MILPs") that seek to minimize the cost of committing to power generation units and allocating output, given a set of constraints. However, the difference here is that future predictions are uncertain, and probabilistic optimization can handle this uncertainty by using a set of future scenarios (e.g., multiple scenarios) instead of a single predictive estimate of load, solar energy availability, price, etc., over time. Each scenario may be a different realization with its own associated probabilities, and the sum of probabilities across all scenarios equals 100%. While both deterministic and probabilistic models are MILPs, probabilistic optimizations are far more difficult to solve due to the increased complexity of the model (curse of dimensionality) and are computationally expensive to do so. However, probabilistic optimizations are superior to deterministic models (i.e., lower power generation costs) because they consider a range of possible future outcomes compared to a single future outcome, since actual results are more likely to fall within a range of outcomes compared to closely matching a single outcome. These advantages of probabilistic optimization motivated the inventors to develop a method to mimic the behavior of probabilistic optimization in a deterministic setting in order to obtain similar results, but much faster in terms of computation time.

[0091] MILP optimization problems can be solved using solvers such as CPLEX, Gurobi, Mosek, Xpress, or other types of solvers. This disclosure does not depend on how those skilled in the computational art ultimately choose to solve MILP.

[0092] The optimized storage schedule may be the output of a probabilistic model. An extended deterministic model may take the predicted storage schedule as input. The predicted storage schedule may closely approximate the optimized storage schedule from the probabilistic model. The storage schedule may be predicted using a gradient-boosted model that minimizes the mean squared error ("MSE"). Similar to the probabilistic model, the extended deterministic model may also output an optimized storage schedule.

[0093] The mean squared error is defined as follows, where Y(i) is the predicted value and Y^(i) is the actual value that the method is trying to predict. This is a commonly used metric used to capture the accuracy of a model.

[0094]

number

[0095] In one embodiment, scenario data from PGscen (a tool made public to support research on probabilistic optimization by a team funded by the DOE ARP-E program) may be used to generate scenarios for the previous day's load and solar energy. At a high level, all scenario generation methods aim to learn uncertainty information in past predictions and then use a model that characterizes this uncertainty to create future scenarios. ML methods for predicting critical constraints are independent of the scenario generation methods. ML methods require only a set of scenarios as input. Furthermore, or alternatively, scenario generation methods developed within the organization may be used.

[0096] Some generator settings may include one or more of the following: minimum uptime, minimum downtime, ramp-up speed, ramp-down speed, startup cost, variable operating and maintenance costs, minimum power generation capacity, and thermal requirements.

[0097] Method 500 may further include the steps of: testing a machine learning model for predicting the required committed capacity of an electric utility by generating a storage schedule for test data; training the machine learning model for predicting the required committed capacity of an electric utility with the test data; predicting the required committed capacity of an electric utility using the machine learning model trained on the test data for predicting the required committed capacity of an electric utility; and using each predicted, required capacity to extend a deterministic output allocation optimization model.

[0098] Referring again to Figure 5, in step 520 of method 500, the controller 112 schedules the power generation equipment, generators, and power generation by the electric utility based on the electric utility's predicted and required committed capacity.

[0099] The method described herein combines the advantages of probabilistic and deterministic methods by applying a machine learning model to the predicted total committed capacity from conventional thermal units calculated using probabilistic optimization to predict the required committed capacity of an electric utility, and uses the predicted, required committed capacity of the electric utility as input for an extended version of a deterministic output allocation optimization model. The method described herein fills the gap between probabilistic and deterministic optimization and applies machine learning to produce high-quality results that are comparable to probabilistic methods but at a speed similar to that of deterministic methods.

[0100] The operating steps of Method 500 are understood to be performed by the computer or processor described herein, which loads and executes software code or instructions, and the instructions are tangibly stored in a tangible, non-temporary, computer-readable storage medium, such as a magnetic medium, such as a computer hard disk; an optical medium, such as an optical disc; a solid-state memory, such as flash memory; or other storage mediums known in the art. Therefore, any of the functions performed by the computer or processor described herein are implemented in software code or instructions tangibly stored in a tangible, non-temporary, computer-readable storage medium. When such software code or instructions are loaded and executed by the computer or processor, the computer or processor may perform any of the functions of the computer or processor described herein, including any step of the method described herein.

[0101] As used herein, the terms “software code” or “code” refer to any instruction or set of instructions that affect the operation of a computer or processor. Any instruction or set of instructions may exist in a computer-executable form, such as machine code, which is a set of instructions and data that is executed directly by the computer’s central processing unit or controller; in a human-readable form, such as source code, which can be compiled for execution by the computer’s central processing unit or controller; or in an intermediate form, such as object code, which is generated by a compiler. As used herein, the terms “software code” or “code” also include any human-readable set of computer instructions or instructions, such as a script, which can be executed immediately with the help of an interpreter executed by the computer’s central processing unit or controller.

[0102] In the above example, the energy management system 101, energy system 102, energy application 110, load and price forecasting system 114, controller 112, etc., each include a network communication interface 206 for wired or wireless communication over one or more networks 118. Network 118 may support data communication by equipment on-premises via wired (e.g., cable or fiber) media, wireless (e.g., Wi-Fi, Bluetooth®, ZigBee, LiFi, IrDA, etc.), or a combination of wired and wireless technologies. The specific design and embodiment of the network communication interface 206 may depend on the communication network on which the computing device 116 is intended to operate. For example, the network communication interface 206 may include a wireless local area network (WLAN) card, an Integrated Digital Network (ISDN) card, a cellular modem, a satellite modem, a modem configured to provide data communication connectivity over the Internet, a network card with an Ethernet port, a device with a radio frequency receiver and transmitter, a device with an optical receiver and transmitter, etc. In some examples, the network communication interface 206 may be designed to operate over the network 118. The network communication interface 206 may be configured to transmit and receive electrical, electromagnetic, or optical signals, which may represent various types.

[0103] Any of the functions of the computer implementation method 500 for predicting the required committed capacity of an electric utility, including the essential capacity control programming 302 and the predictive storage control programming 326 described herein, for energy management systems 101, energy systems 102, power applications 110, load and price forecasting systems 114, controllers 112, etc., can be embodied in one or more applications or firmware as described above. According to some embodiments, “function”, “functions”, “application”, “applications”, “instruction”, “instructions”, or “programming” is a program(s) that performs functions defined within the program. Various programming languages ​​can be used to create one or more applications structured in various ways, such as object-oriented programming languages ​​(e.g., Objective-C, Java, or C++) or procedural programming languages ​​(e.g., C or assembly language).

[0104] In the examples above, the energy management system 101, energy system 102, power application 110, load and price forecasting system 114, controller 112, etc., may each include a processor. As used herein, a processor 202 is a hardware circuit having elements structured and arranged to perform one or more processing functions, typically various data processing functions. A processor 202 may provide the ability to run, control, operate, or store multiple processes, applications, or programs. In some examples, a processor 202 may be configured to provide parallel processing capabilities, enabling devices associated with the processor 202 to run multiple processes simultaneously. In some examples, a processor 202 may be comprised of virtualization technology. Other types of processor configurations may be implemented to provide the functions described herein.

[0105] Memory 204 may include a non-temporary computer-readable medium that, when executed by at least one processor (e.g., processor 202), can store instructions causing at least one processor 202 to execute one or more processes described herein. The non-temporary computer-readable medium may include any type of physical memory that can store information or data readable by at least one processor. The non-temporary computer-readable medium may include, for example, random access memory (RAM), read-only memory (ROM), compact disc read-only memory (CD-ROM), digital versatile disc (DVD), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), non-volatile random access memory (NVRAM), volatile memory, non-volatile memory, hard drives, flash drives, disks, caches, registers, optical data storage media, patterned physical media, or networked versions thereof. The non-temporary computer-readable medium may include multiple structures that may be located at local or remote locations.

[0106] Memory 204 may include flash memory (non-volatile or persistent memory), read-only memory (ROM), and random access memory (RAM) (volatile memory). RAM functions as short-term storage for instructions and data being processed by processor 202, for example, as working data processing memory. Flash memory typically provides longer-term storage.

[0107] Needless to say, other storage devices or configurations may be added to or replace the storage devices or configurations in the examples. Such other storage devices may be implemented using any type of storage medium having computer or processor-readable instructions or programming stored therein, and may include, for example, tangible memory of a computer, processor, or any or all of related modules.

[0108] Therefore, machine-readable or computer-readable media can take the form of many tangible storage media. Non-volatile storage media include, for example, optical disks or magnetic disks, such as any storage device in any computer(s), which may be used to implement client devices, media gateways, transcoders, etc., as shown in the drawings. Volatile storage media include dynamic memory, such as the main memory of such a computer platform. Tangible transmission media include coaxial cables, copper wires, and optical fibers, including wires, which include buses in computer systems. Carrier transmission media may take the form of acoustic waves or optical waves, such as electrical signals or electromagnetic signals, or those generated during radio frequency (RF) and infrared (IR) data communications. Therefore, common forms of computer-readable media include, for example, floppy disks, flexible disks, hard disks, magnetic tapes, any other magnetic media, CD-ROMs, DVDs or DVD-ROMs, any other optical media, punched card paper tapes, any other physical storage media having a pattern of holes, RAM, PROMs and EPROMs, FLASH-EPROMs, any other memory chips or cartridges, carriers for transporting data or instructions, cables or links for transporting such carriers, or any other media from which a computer can read programming code and / or data. Many of these forms of computer-readable media can be involved in transporting one or more sequences of one or more instructions to a processor for execution.

[0109] According to exemplary embodiments of this disclosure, one or more processors and control circuits may include one or more of any known general-purpose processors or integrated circuits, such as a central processing unit (CPU), a microprocessor, a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a digital signal processor (DSP), or, as desired, another suitable programmable processing device or computing device or circuit specifically programmed to perform operations to achieve the results of the exemplary embodiments described herein. The processor(s) may include and be configured to perform functions of exemplary embodiments of this disclosure, such as a computer implementation method 400 for predicting the required committed capacity of an electric utility. The functions may be performed as desired via program code encoded or recorded on the processor(s) or stored in read-only memory (ROM), erasable programmable read-only memory (EPROM), or other non-volatile memory devices. Thus, such computer programs may represent controllers of computing devices.

[0110] In another exemplary embodiment, program code such as a computer implementation method 400 for predicting the required committed capacity of an electric utility may be provided in a computer program product that has a non-temporary computer-readable medium such as a magnetic storage medium (e.g., a hard disk, floppy disk, or magnetic tape), an optical medium (e.g., any type of compact disc (CD) or any type of digital video disc (DVD), or any other compatible non-volatile memory device as desired), and which is downloaded to the processor(s) for execution as desired once the non-temporary computer-readable medium is positioned in communicative contact with the processor(s).

[0111] One or more processors 202 may be included as desired in a computing system comprising components such as memory, hard drives, input / output (I / O) interfaces, communication interfaces, displays, and any other suitable components. The exemplary computing device 116 may also include a communication interface 206. The communication interface 206 may be configured to enable the transfer of software and data between the computing device 116 and external devices. The exemplary communication interface may, as desired, include a modem, a network interface (e.g., an Ethernet card), a communication port, a PCMIA slot and card, or any other suitable network communication interface. The software and data transferred via the communication interface 206 may be in the form of signals, which may be electronic signals, electromagnetic signals, optical signals, or other signals, as will be apparent to those skilled in the art in the relevant field. The signals may travel through a communication path, which may be configured to carry the signals, and may be implemented using wires, cables, optical fibers, telephone lines, cellular telephone lines, radio frequency links, or any other suitable communication links.

[0112] If this disclosure is implemented using programming or software, including a computer implementation method 400 for predicting the required committed capacity of an electric utility, the programming or software may be stored in a computer program product or a non-temporary computer-readable medium and loaded onto a computing device using a removable storage drive or a communication interface. In exemplary embodiments, any computing device such as the controller 112 disclosed herein may also include a display interface that outputs display signals to a display device, such as an LCD screen, plasma screen, LED screen, DLP screen, CRT screen, or any other suitable graphical interface, as desired.

[0113] The terms and expressions used herein shall be understood to have the ordinary meanings given to such terms and expressions in relation to the respective areas of the corresponding investigations and studies, unless a specific meaning is otherwise explained herein. Relative terms such as "first" and "second" are used solely to distinguish one entity or action from another entity or action and do not necessarily imply or require an actual relationship or order between such entities or actions. The terms "comprises," "comprising," "includes," "including," "has," "having," "containing," "contains," "together," "formed from," or any other variation thereof are intended to encompass non-exclusive inclusion, such as including a list of elements or steps, or including a process, method, article, or apparatus that includes such elements or steps, rather than including only those elements or steps, and may also include other elements or steps that are not explicitly listed or that are inherent to such process, method, article, or apparatus. An element preceded by "a" or "an" does not, unless otherwise specified, exclude the presence of additional identical elements in a process, method, article, or apparatus that includes that element. Unless otherwise specified, the article "a" or "an" preceding an element refers to one or more of those elements.

[0114] Unless otherwise specified, all measurements, values, ratings, locations, sizes, angles, and other designations described herein, including in the following claims, are approximate and not strict. Such quantities are intended to have a reasonable range that is consistent with the function to which they relate and with what is customary in the art to which they relate. For example, unless expressly specified otherwise, parameter values, etc., may vary by approximately ±5% or ±10% from the stated quantities. The terms “approximately” and “substantially” mean that parameter values, etc., may vary by up to ±10% from the stated quantities.

[0115] Furthermore, in the modes for carrying out the invention described above, it can be understood that various features are grouped together in various examples in order to rationalize this disclosure. This method of disclosure should not be interpreted as reflecting an intention that the claimed examples require more features than are explicitly stated in each claim. Rather, as reflected in the following claims, the subject matter to be protected is not all of the features of any single disclosed example. Thus, the following claims are incorporated here into the modes for carrying out the invention, and each claim is based on itself as separately claimed subject matter.

[0116] While the best forms and / or other examples have been described above, it will be understood that various modifications are possible, that the subject matter disclosed herein can be implemented in various forms and examples, and that these may apply to numerous applications, of which only a portion are described herein. The following claims are intended to claim all possible modifications and variations that fall within the true scope of the Concept.

[0117] The scope of protection is limited solely by the claims that follow this specification, which, when interpreted in light of this specification and the following application history, is intended and should be interpreted to encompass all structural and functional equivalents, insofar as it is consistent with the ordinary meaning of the language used in the claims. Nevertheless, none of the claims are intended, nor should they be interpreted, to encompass subject matter that does not meet the requirements of Sections 101, 102, or 103 of the Patent Act. Any unintended inclusion of such subject matter is disallowed herein.

Claims

1. A computer implementation method for predicting the required guaranteed capacity of an electric utility, (a) Performing probabilistic optimization of raw data in order to generate the total guaranteed capacity from conventional thermal units as target data, wherein the raw data includes power grid operating conditions, (b) In order to generate training data, the total guaranteed capacity from the conventional thermal unit is combined with raw features and designed features, (c) Using the generated training data, train a machine learning model to predict the required committed capacity of the electric utility; (d) Predicting the required committed capacity of the electricity utility using the trained machine learning model for predicting the required committed capacity of the electricity utility, (e) To run an extended version of the deterministic power allocation optimization model based on the predicted required committed capacity of the electric utility, The computer implementation method, including the above.

2. The computer implementation method according to claim 1, wherein the raw data includes solar power generation or wind power generation, weather forecasts, multiple sets of forecasts of net load demand per hour, a net exchange schedule, a region, and a net storage schedule.

3. The computer implementation method according to claim 2, wherein the net storage schedule includes an optimized schedule and a predicted schedule.

4. The computer implementation method according to claim 2, wherein the raw data further includes the total capacity of the time-based load-following startup delay type units, which is determined by the probabilistic optimization.

5. The computer implementation method according to claim 1, wherein training the machine learning model for predicting the required committed capacity of the electric utility comprises designing additional input features from the raw data and selecting the additional input features and a subset of features from the raw data.

6. The computer implementation method according to claim 5, wherein the subset of features includes average net demand, delayed average net demand, average net demand / average total demand, month within the year, Is_weekday, net storage schedule, and net exchange schedule.

7. The computer implementation method according to claim 5, wherein the selection of the subset of features includes measuring the dependency between each feature and a target variable representing the total conventional capacity expected by the probabilistic optimization, ensuring that the dependency is equal to zero when the two random variables are independent, capturing nonlinear dependencies, removing low-information features, reducing highly correlated features, performing a Lasso regression, a Ridge regression, and selecting top features having the best mutual information score with the target data.

8. The computer implementation method according to claim 1, wherein training the machine learning model includes a neural network (PyTorch), a random forest (Sklearn), and gradient boosting (LightGBM and XGBoost).

9. The computer implementation method according to claim 1, further comprising training a machine learning model to predict optimal storage behavior by using observed past storage states.

10. The computer implementation method according to claim 9, wherein the predicted storage is used as an input feature for the machine learning model for predicting the required committed capacity of the electric utility.

11. The computer implementation method according to claim 9, wherein the machine learning model for predicting the optimal storage behavior generates a storage schedule for test data by training the machine learning model for predicting the required guaranteed capacity of the electric utility to predict the charging and discharging behavior of the storage.

12. The computer implementation method according to claim 9, wherein the machine learning model for predicting the optimal storage behavior is trained daily to generate predictions of storage behavior, and the predictions of storage behavior are used as input features to the machine learning model for predicting the required committed capacity of the electric utility.

13. The computer implementation method according to claim 1, wherein the machine learning model for predicting the required committed capacity of the electric utility is used daily to generate capacity constraints for the extended deterministic output allocation optimization model, which is improved by retraining the machine learning model for predicting the required committed capacity of the electric utility with updated data.

14. A computer implementation method according to claim 1, further comprising: testing the machine learning model for predicting the required committed capacity of the electric utility by generating a storage schedule for test data; training the machine learning model for predicting the required committed capacity of the electric utility with the test data; predicting the required committed capacity of the electric utility using the machine learning model trained on the test data for predicting the required committed capacity of the electric utility; and using the predicted required committed capacity to extend the deterministic output allocation optimization model.

15. The computer implementation method according to claim 14, wherein the predicted required guaranteed capacity is used to extend the deterministic output allocation optimization model by running the deterministic output allocation optimization model using the capacity prediction as the minimum constraint on the conventional capacity in the previous day ("DA") cycle.

16. The computer implementation method according to claim 1, wherein predicting the capacity of the aforementioned electric public utility predicts the total heat guaranteed capacity value at predetermined time intervals.

17. The computer implementation method according to claim 16, wherein the total heat capacity value includes the guaranteed capacity of the thermal power smoke generating equipment.

18. The computer implementation method according to claim 17, wherein the thermal power generation equipment includes a nuclear power plant, a coal plant, or a natural gas plant, a combined cycle ("CC") source, a combustion turbine ("CT"), or a steam turbine ("ST").

19. The computer implementation method according to claim 16, wherein predicting the required guaranteed capacity of the aforementioned electric utility business is performed by using the power grid operating conditions to predict the total heat guaranteed capacity value at predetermined time intervals.

20. The computer implementation method according to claim 1, wherein the power grid operating conditions include net demand per hour and net exchange.

21. The computer implementation method according to claim 1, wherein the target data includes the total guaranteed capacity of the conventional thermal units per hour, which is scheduled by the probabilistic optimization in the previous day ("DA") cycle.

22. The computer implementation method according to claim 1, wherein the extended version of the deterministic output allocation optimization model is run daily to produce unit commitments and output allocation plans for a customer, and the extended version of the deterministic output allocation optimization model uses the predicted, required committed capacity of the electric utility as input.

23. The computer implementation method according to claim 22, wherein the extended version of the deterministic optimization model outputs the optimized storage schedule, and the probabilistic optimization uses the optimized storage schedule as an input feature.

24. The computer implementation method according to claim 1, further comprising the electric utility scheduling the on and off of power generation equipment, generators, and power generation based on the predicted and required committed capacity of the electric utility.

25. A non-temporary computer-readable medium containing programming, wherein a processor executes the programming, and a computing device (a) Performing probabilistic optimization of raw data in order to generate the total guaranteed capacity from conventional thermal units as target data, wherein the raw data includes power grid operating conditions, (b) In order to generate training data, the total guaranteed capacity from the conventional thermal unit is combined with the raw features and the designed features, (c) Using the generated training data, train a machine learning model to predict the required committed capacity of the electric utility; (d) Using the trained machine learning model, predict the required committed capacity of the electric utility, (e) To run an extended version of the deterministic power allocation optimization model based on the predicted required committed capacity of the electric utility, The non-temporary computer-readable medium configured to perform the following actions.

26. A computing device, Memory and A processor coupled to the aforementioned memory, The programming in the memory, wherein the processor executes the programming, and the computing device, (a) Performing probabilistic optimization of raw data in order to generate the total guaranteed capacity from conventional thermal units as target data, wherein the raw data includes power grid operating conditions, (b) In order to generate training data, the total guaranteed capacity from the conventional thermal unit is combined with the raw features and the designed features, (c) Using the generated training data, train a machine learning model to predict the required committed capacity of the electric utility; (d) Using the trained machine learning model, predict the required committed capacity of the electric utility, (e) To run an extended version of the deterministic power allocation optimization model based on the predicted required committed capacity of the electric utility, The programming is configured to perform the following: The computing device comprising the above.

Images