WO2019117957A1

WO2019117957A1 - System device, and method for mode-based energy storage management using machine learning

Info

Publication number: WO2019117957A1
Application number: PCT/US2017/066778
Authority: WO
Inventors: Gonzague HENRI; Carlos Carrejo; Ning Lu
Original assignee: Total Solar International; North Carolina State University
Priority date: 2017-12-15
Filing date: 2017-12-15
Publication date: 2019-06-20

Abstract

A method, system, and device for controlling an energy storage device are provided. The method includes identifying a predetermined number of features based on forecasted data and data obtained from the energy storage device. The method also includes identifying an operating mode from a plurality of operating modes for a predetermined operating period using a machine learning model and operating the energy storage device in the operating mode for the predetermined operating period.

Description

SYSTEM, DEVICE, AND METHOD FOR MODE-BASED ENERGY STORAGE

MANAGEMENT USING MACHINE LEARNING

BACKGROUND

[0001] High penetration of residential and commercial rooftop photovoltaic (PV) systems may increase power fluctuations in distribution feeders and reverse power flow direction. As a result, utilities start to experience many voltage issues, such as over-voltage, large voltage ramps, and voltage swings. The intermittency of the solar generation resources and backfeeding power are the two main causes of these problems. To alleviate these issues, utilities are revising their rate structure to provide incentive for the customers to self-consume the solar energy they generated and limit the amount of power that can be backfed to the main grid. For example, Germany lowered the feed-in-tariff as described in P. Denholm and R. Margolis,“Energy storage requirements for achieving 50% solar photovoltaic energy penetration in California,” Nrel, August 2016, while in Hawaii, backfeeding is no longer allowed for newly installed PV systems, per public utility commission order in 2014.

Therefore, installing energy storage devices (ESDs) to store excess solar power and smooth the power fluctuations is an increasingly attractive option for residential and commercial PV systems.

[0002] For residential PV applications, battery storage systems are used to fulfill different objectives. The objectives of scheduling and dispatching the battery systems include minimizing utility bills as described in T. Hubert and S. Grijalva,“Modeling for residential electricity optimization in dynamic pricing environments,” IEEE Transactions on Smart Grid, vol. 3, no. 4, pp. 2224-2231, 2012, smoothing PV outputs, maximizing self-consumed solar energy as described in Y. Wang, X. Lin, and M. Pedram,“A near-optimal model-based control algorithm for households equipped with residential photovoltaic power generation and energy storage systems,” IEEE Transactions on Sustainable Energy, vol. 7, no. 1, 2016, or providing different grid services as described in M. Giuntoli and D. Poll,“Optimized thermal and electrical scheduling of a large scale virtual power plant in the presence of energy storages,” IEEE Transactions on Smart Grid, vol. 4, no. 2, pp. 942-955, 2013. A variety of optimization methods to solve those scheduling problems and dispatch the battery power outputs in real-time, such as dynamic programming, fuzzy logic, mixed integer programming (MIP), and stochastic programming are described in L. Liu, Y. Zhou, Y. Liu, and S. Hu,“Dynamic programming based game theoretic algorithm for economical multi- user smart home scheduling,” Midwest Symposium on Circuits and Systems, pp. 362-365, 2014, Zhi Wu, Xiao-Ping Zhang, J. Brandt, Su-Yang Zhou, and Jia-Ning Li,“Three control approaches for optimized energy flow with home energy management system,” IEEE Power and Energy Technology Systems Journal, vol. 2, no. 1, pp. 21-31, 2015, M. C. Bozchalui, S. A. Hashmi, H. Hassen, C. A. Canizares, and K. Bhattacharya,“Optimal operation of residential energy hubs in smart grids,” IEEE Transactions on Smart Grid, 2012, and Z. Yu,

L. Jia, M. C. Murphy-Hoye, A. Pratt, and L. Tong,“Modeling and stochastic control for home energy management,” IEEE Transactions on Smart Grid, vol. 4, no. 4, pp. 2244-2255, 2013, respectively. Those methods try to find the optimal power outputs of the battery at each dispatch interval over a given scheduling period and meet the battery operational constraints.

[0003] For residential applications, the main issue is to find when to charge and when to discharge instead of the optimal charging and discharging power. Accordingly, what is needed, as recognized by the present inventors, is a method for controlling the operation of an energy storage device. [0004] The foregoing“Background” description is for the purpose of generally presenting the context of the disclosure. Work of the inventor, to the extent it is described in this background section, as well as aspects of the description which may not otherwise qualify as prior art at the time of filing, are neither expressly or impliedly admitted as prior art against the present invention.

SUMMARY

[0005] The present disclosure relates to a method that identifies, using processing circuitry, identifying, using processing circuitry, a predetermined number of features based on forecasted data and data obtained from the energy storage device and identifies an operating mode. The operating mode is identified from a plurality of operating modes for a

predetermined operating period using a machine learning model. The processing circuitry also operates the energy storage device in the operating mode for the predetermined operating period.

[0006] The present disclosure also relates to an energy storage device that includes a controller. The controller is configured to identify a predetermined number of features based on forecasted data and data obtained from the energy storage device, identify an operating mode from a plurality of operating modes for a predetermined operating period using a machine learning model, and operate the energy storage device in the operating mode for the predetermined operating period.

[0007] The present disclosure also relates to a system that includes an energy storage device and a controller. The energy storage device is configured to operate in an operating mode indicated by a control signal received from a controller. The controller is configured to identify a predetermined number of features based on forecasted data and data obtained from the energy storage device, identify an operating mode from a plurality of operating modes for a predetermined operating period using a machine learning model, and operate the energy storage device in the operating mode for the predetermined operating period.

[0008] The foregoing paragraphs have been provided by way of general introduction, and are not intended to limit the scope of the following claims. The described embodiments, together with further advantages, will be best understood by reference to the following detailed description taken in conjunction with the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

[0009] A more complete appreciation of the disclosure and many of the attendant advantages thereof will be readily obtained as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings, wherein:

[0010] FIG. 1 is an exemplary diagram of an energy storage management system according to one example;

[0011] FIG. 2 is a flowchart that shows a mode selecting process according to one example;

[0012] FIG. 3 is a flowchart that shows a method for operating an energy storage device (ESD) according to one example;

[0013] FIG. 4 is a schematic that shows exemplary results for a plurality of machine learning functions as a function of the training length according to one example;

[0014] FIG. 5 is a schematic that shows exemplary results for the plurality of machine learning functions as a function of the number of features in a training set;

[0015] FIG. 6 is a schematic that shows the accuracy for each mode according to one example; [0016] FIG. 7 is a schematic that shows the energy cost for machine learning mode based control and for economic model predictive control (EMPC) mode based control; and

[0017] FIG. 8 is an exemplary block diagram of a computer according to one example.

DETAILED DESCRIPTION

[0018] Referring now to the drawings, wherein like reference numerals designate identical or corresponding parts throughout several views, the following description relates to a system, a device, and associated methodology for controlling an energy storage system. Charge and discharge modes are identified so that an overall performance may be optimized in the presence of load forecasting errors using a mode selection process. The method described herein may be used to control energy storage devices (ESDs). The method described herein learns from past operation history to capture a relationship between operational conditions and the mode selection process.

[0019] FIG. 1 is an exemplary diagram of an energy storage management system 100. The system may include an energy storage device (ESD) 102. The ESD 102 stores renewable energy generated by a renewable generation system 108 such as solar photovoltaics (PVs). The renewable generation system 108 may include a wind power generation system (e.g., wind mills), a hydraulic energy source, a micro combined heat and power (CHP) unit for heating and electricity generation, or any other energy system from renewable resources such as rain, tides, or waves. The ESD 102 may be an electrical energy storage device, a fuel cell, a thermal energy storage device, a bioelectrochemical energy storage device, a hybrid energy storage device, or the like. The ESD 102 may also store energy supplied by a grid 116.

[0020] In one implementation, the ESD 102 may be a part of a home energy management system. The ESD 102 may be used to supply power to a load 106 when the cost of electricity is high and be recharged when the cost is low. The ESD 102 may be a part of a building energy management system at a commercial building, a residential building, or an industrial building. The grid 116 may supply energy to the load 106 when the cost of electricity is low.

[0021] In one implementation, the method described herein may be used in a power microgrid system that includes renewable energy sources and at least the ESD 102. Further, the microgrid system may be a hybrid microgrid that includes renewable energy sources, the ESD 102, and a second energy source such as a diesel/gas generator.

[0022] In one implementation, the method described herein may be applied in aggregated distributed energy resources systems that include electric battery storage in a commercial building, an industrial building, or a residential building or a home.

[0023] In one implementation, the ESD 102 may be a part of an aerospace structure such as a satellite, an aircraft, a spacecraft, and other space vehicles.

[0024] In one implementation, the ESD 102 may include a battery in an electric vehicle. In another example, the ESD 102 may be a large-capacity battery bank used in a data center or a smart grid.

[0025] The ESD 102 is controlled via a controller 104. The controller 104 may determine a real time mode of operation for the ESD 102 using a machine learning (ML) based control process based on a forecasted load profile and renewable energy production information. The operating mode may be identified from a plurality of operating modes. The information may be obtained using an information acquisition module 110 that connects to a database 112 and/or input sources 114. The database 112 may store forecast models, operating conditions associated with the plurality of operating modes, or other information used by the controller 104 to determine an operating mode. The database 112 may also store historical data and the associated operating mode. The input sources 114 may include weather websites, energy supplier websites, or other sources that may have real-time weather data. The information acquisition module 110 may connect to the input sources 114 via a network. Suitable networks can include or interface with any one or more of a local intranet, a Personal Area Network (PAN), a Local Area Network (LAN), a Wide Area Network (WAN), a

Metropolitan Area Network (MAN), a Virtual Private Network (VPN), or a storage area network (SAN).

[0026] The modules described herein may be implemented as either software and/or hardware modules and may be stored in any type of computer-readable medium or other computer storage device. For example, each of the modules described herein may be implemented in circuitry that is programmable (e.g., microprocessor-based circuits) or dedicated circuits such as application specific integrated circuits (ASICS) or field

programmable gate arrays (FPGAS). In one embodiment, a central processing unit (CPU) could execute software to perform the functions attributable to each of the modules described herein. The CPU may execute software instructions written in a programing language such as Java, C, or assembly. One or more software instructions in the modules may be embedded in firmware, such as an erasable programmable read-only memory (EPROM).

[0027] In some implementations, the processes associated with each of the modules may be performed by one or more processors of the controller 104 or other computing resources, which can include cloud computing resources.

[0028] For example, the processes may be performed by a computer 826 shown in FIG. 8. The computer 826 may include a CPU 800 and a memory 802 as shown in FIG. 8. The CPU 800 may execute software instructions written in a programing language such as Java, C, or assembly. In one implementation, the database 112 may be implemented in the memory 802 of the computer 826. In one implementation, the computer 826 may control a plurality of energy storage devices. The plurality of energy storage devices may be associated with a single entity (e.g., a home) or a plurality of entities (e.g., a community). For example, the computer 826 may identify an operating schedule and operating modes for each of the plurality of entities based on one renewable energy forecast when the entities are located in a geographical area that has common weather data.

[0029] In one implementation, five real-time operations modes may be used as shown in Table 1.

Table 1 : Simplified modes of the ESD controller

[0030] The ESD power output in each mode is calculated as follows. First, the controller 104 may calculate the netload, P_net , as

where is the load of the household, Psol is the renewable generation system power output (e.g., solar power output) and the ESD energy limit (or battery energy limit),

where is the charging power cap, E_B is the current ESD capacity, and is the time

interval for mode-based operation (minute).

[0031] For the ESD discharging power cap, for time interval i based on the current

ESD energy level, and the ESD energy limit,

where is the minimum ESD capacity. [0032] The power charging cap shows that, considering how much more energy can be stored in the ESD, what the maximum charging power of the ESD is for the

interval. Then, the ESD charging/discharging power at the

time interval,

at mode m can be formulated as

where P_rated is the rated battery power and N is the number of time intervals.

[0033] Next, the model selection process for selection of the right amount of features for the machine learning process, the minimum amount of training data, and the ML algorithm architecture in the case of Neural Networks (NN) are described.

[0034] In one implementation, data from the PECAN street project have been used. 149 houses are selected with PV installations. The 149 houses are located in Austin, Texas and have 8760 hourly points collected in 2015. For each house, the mode-based control algorithm using the actual load and PV data (i.e. the forecast is perfect) is implemented to obtain the optimal operation modes, for hour i house /. This process gives the

optimal modes at 8760 hours for all the 149 houses.

[0035] The data generated is then used to train different ML algorithms. For dispatch interval i, the mode-based algorithm with an average load forecast is run. Thus, two benchmarks are obtained: modes selected using a perfect load forecaster and modes selected using an average load forecaster. The perfect forecast vector may be used to calculate the mode selection accuracy, the vector based on the average load forecast may be used as a benchmark to demonstrate the improvement of the ML approach compared with the economic model predictive control (EMPC) based approach as described later herein.

[0036] At each time step, the following 17 variables are recorded: the state of charge (SOC), PV and load from last time step

forecast for the current time step the remaining energy storage capacity to charge or discharge ,

the sum of forecasted PV for the next 24 hours , temperature, hour, day of the

week, month, day of the month, weekday or a weekend, and the prices to import electricity

Thus, the training set may contain 14 features in total.

[0037] The first step is to identify what features among the set previously described have an impact on the mode selection process. To achieve this goal, multiple analysis of variance (ANOVA) tests are performed to select the K-best features as described in I. Guy on, A. Elisseeff, and A. M. De,“An introduction to variable and feature Selection,” Journal of Machine Learning Research, vol. 3, pp. 1157-1182, 2003. Note that where

From the ANOVA results, the most important features are identified. The 12

months data sets are divided into training and testing sets such that the training set contains 11 months of data and the testing set contains 1 month of data.

[0038] The second step is to determine the best number of features for achieving the highest accuracy in the mode prediction. To do so, ML algorithms are trained using different number of K features following the order of importance. Then, each algorithm is tested on the same test set to find the best performing one.

[0039] The third step is to quantify the impact of the training set length on the accuracy. The accuracy may be first tested using 1 month training data and then the length of the training data is increased from 1 to 11 months with an increment of 1 month. The ML algorithm achieving the best overall performance is selected as the algorithm to serve for the control. Then, the ML algorithm accuracy is tested for each mode in order to verify that the increase in accuracy is not just for any specific mode. Following the method described in S.

Kotsiantis, I. Zaharakis, and P. Pintelas,“Supervised machine learning: A review of classification techniques,” Informatica, vol. 31, pp. 249-268, 2007, with continuous features and a large data set, support vector machine (SVM) and NN perform the best. The selection of the best learning architecture is obtained after a sensitivity analysis on the feature scenarios and NN architectures. Several architectures are listed in Table 2. The algorithm with the best performance is selected.

Table 2: List of the different algorithms implemented

[0040] Next, the ML algorithm for feature selection and mode prediction is described. The method for the real-time use of a ML algorithm for residential ESD control may be divided into two stages. The first stage acquires the needed data for the training and train the model in an offline fashion. The second stage uses the NN to select the mode for the next time step.

[0041] A mode-based control algorithm with average load forecast can be used to operate the system until the recorded data is sufficient to train the machine learning algorithm for mode prediction. For example, the method described in G. Henri, N. Lu, and C. Carrejo,“Design of a Novel Mode-based Energy Storage Controller for Residential PV Systems,” in Proc. of 2017 Power & Energy Society ISGT Europe Conference, Turino, 2017, may be implemented. The inputs at this stage are the electricity prices, the temperature, the load, and the PV production. This task may last until enough data has been acquired to perform the offline training. If the customer has enough historical data from his smart meter, then the offline training can start without the data collection period.

[0042] In one implementation, the EMPC controller is a Mixed Integer Programming (MIP) based scheduling algorithm. In a first stage, a 24-hour schedule is determined (e.g., by the controller 104) based on the 24-hour ahead forecast of PV, load, and electricity price. The 24 hour schedule reflects the charging and discharging power of the ESD 102 for optimizing the energy bill for the predetermined period (e.g., next 24-hour period) without considering the feasibility of the ESD control actions. Then, control modes that are unfit are eliminated based on the required actions that the ESD 102 is identified to take for the next hour or other period. For example, if the EMPC results show that for the next hour, the battery should charge at 2 kW, then all the discharging modes are eliminated and all the charging modes are selected for the next stage comparison. Note that although in this case, an EMPC based approach is used to minimize the bill for the user over the 24-hour period, other methods may be implemented for obtaining the operational trends. At the end of each first stage, the optimization parameter (i.e., optimization variable or value, total cost) for the next scheduling horizon and the optimal ESD output for the next time step are obtained by the controller 104. The optimization parameter includes the cost of import and export energy as well as the cost for battery degradation. The feasible modes are also obtained.

[0043] The inputs of a second stage include: the feasible mode, the forecasted data, the optimization parameter for the next day, and the control actions for the next hour. First, the optimization parameter for the next hour along with the SOC at the end of the next time step is calculated by the controller 104, for each feasible mode. Then, for each feasible mode, an optimization is run starting at the end of the next time step. The horizon of the optimization is reduced by one time step by the controller 104 (e.g., by one hour). The optimization may start after the next time step with the newly calculated SOC. Then, the result of the optimization is summed with the optimization parameter of the next hour. The total optimization parameter for the next day for each feasible mode is determined by the controller 104 as described later herein. The mode with the least optimization parameter is selected. If the optimization parameter found in the first stage is lower, a charging or discharging limit may be applied. The limit corresponds to the ESD output found in the first stage.

[0044] A perfect load forecast, a neural network forecaster, and the average load as the load forecast may be used to forecast residential loads. A perfect load forecaster uses actual loads assuming zero forecasting error is used. A neural network forecaster may consist of a Narnet with 25 neurons in the first layer and 10 in the second layer. A load forecaster that forecasts the average load for the whole day is used such that an expected load to solar output ratio is obtained for determining whether the energy storage should charge or discharge.

[0045] Once the historical data are obtained and stored in the memory 802 or database 112, a mode-based algorithm with a perfect forecaster may be run over the historical data set to obtain the optimal mode and battery features (i.e., SOC,

and the forecaster features at each time step. This process is described in

Algorithm 1.

[0046] After the optimal modes for each time step is obtained for the historical data set, an offline training of the ML algorithm is implemented. First, the training data may be regularized; then, the training may be performed on the regularized training set. The regularization parameters as well as the trained ML algorithm may be transmitted to the controller 104 from the server 118. In one implementation, the controller 104 may regularize the training data and train the ML algorithm. The regularization parameters may be used on the testing set.

[0047] Once the training on the historical data is done, the ML algorithm may replace the mode-based control algorithm. The controller 104 may keep updating the model every few days or weeks, use the ML algorithm to predict the mode for the next time step, and operate the ESD based on this mode. This process is illustrated in Algorithm 2.

[0048] NN may be trained using Adam algorithm described in D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,”Iclr, pp. 1-15, 2015. Adam implements a gradient descent algorithm with an adaptive learning rate. Adam algorithm can be used for both offline and online training. It is particularly effective for large data sets. In other implementations, other optimization methods such as hill climbing, genetic algorithms, evolutionary algorithms, simulated annealing, or the like may be implemented.

[0049] To keep updating the ML algorithm, the load, PV, date, and temperature may be recorded at each time step. Then, as described previously herein, a mode-based control algorithm with a perfect forecaster and an equivalent ESD model may be run on the data to find the optimal mode at each time step, and populate the training sets. Once the training set is created, the ML algorithm can be updated.

[0050] For real-time control, the controller 104 may generate at each time step a X_test vector of a length equal to the number of features corresponding to the data used to predict the mode for the next time step. To populate this vector, the controller 104 may obtain data from a smart meter, the ESD, and different application programing interfaces (e.g., temperature, PV forecast, and electricity prices). The vector may be sent to the ML algorithm in order to predict the mode for the next time step. Then, the mode is sent to the ESD for the next time step operation. This set of actions will be repeated at each time step.

[0051] As described previously herein, once the first training on historical data has been done, the ML algorithm is used instead of the mode-based control algorithm. The controller 104 may update the model every predetermined number of days (e.g., 10 days, 30 days, 90 days).

[0052] FIG. 2 is a flowchart that shows a mode selecting process according to one example. At step 202, the controller 104 may check whether the training data was not updated within the predetermined number of days. In response to determining that the training data was not updated within the predetermined number of days, the controller 104 may perform online training as described previously herein at step 204. In response to determining that the training have been updated within the predetermined number of days, the process proceeds to step 206. At step 206, the controller 104 may retrieve and/or determine the inputs (features) of the neural network. For example, the price and PV forecast may be computed. At step 208, the inputs are stored in the training database. At step 210, the controller 104 determines the operation mode using the neural network. At step 212, the identified operation mode is output to plant 212 (e.g., ESD 102).

[0053] FIG. 3 is a flowchart that shows a method for operating an energy storage device (ESD) according to one example. At step 302, the processor or controller 104 may identify a predetermined number of features based on forecasted data and data obtained from the energy storage device as described previously herein.

[0054] At step 304, the controller 104 may identify an operating mode from a plurality of operating modes for a predetermined operating period based on a machine learning model.

At step 306, the controller 104 operates the energy storage device in the operating mode for the predetermined operating period. [0055] To illustrate the capabilities of the energy storage management system described herein, exemplary results are presented.

[0056] The load data used in the simulation is from the Pecan Street data set. 149 houses located in Austin with PV installations are selected. There are 8760 data points for the year 2015. The characteristics of the data set are summarized in Table 3. The electricity tariff is from HECO, as described in Table 4. This tariff encourages customers to self-consume their solar generation and does not valorize backfeed to the grid 116. The ESD used is assumed to be the same for all houses. The ESD is 7kWh@3.3kW with a round-trip efficiency of 90%.

Table 3: Statistical description of 190 houses selected

Table 4: Time-of-use rate in HECO (Hawaii utility)

[0057] Two metrics are used to evaluate the experimental results. The first metric is the accuracy of the mode selection length that represents the test duration, in this case 30 days, and 720 hours. M_predicted represents the modes predicted by the ML algorithm and M_optimai represents the optimal modes.

For each time step i:

[0058] The second metric is a performance index, PMSA, the percentage of the maximum savings achieved for each house is used, described by:

[0059] represents the base cost or the base optimization parameter with no ESD,

is the optimal optimization parameter (i.e., the case with EMPC using a perfect forecast), and is the optimization parameter obtained using one of two algorithms: EMPC + average

load forecast, and ML-based algorithm. The PMSA is calculated for each house in the dataset.

[0060] In one implementation, the learning algorithms are simulated using the Scikit-Learn libray in Python, described in F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B.

Thirion, O. Grisel, M. Blondel, G. Louppe, P. Prettenhofer, R. Weiss, V. Dubourg, J.

Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and Duchesnay,“Scikit-learn: Machine Learning in Python,” vol. 12, pp. 2825-2830, 2012. The optimization problem may be formulated using the Pyomo library and solved by GLPK described in W. E. Hart,“Python optimization modeling objects (Pyomo),” Operations Research/ Computer Science Interfaces Series, vol. 47, pp. 3-19, 2009.

[0061] The ANOVA test has been run on each house. Table 5 represents the counts of features that were among the K best. If a feature is among the K best in one house, the feature is considered as one count. The test is performed with K being comprised between 1 and 14. Thus, every row in table 5 represents the K best features, and each the cell represents the count of the feature as being in the K best features. The count is determined with the ANOVA test in order to find the features that have the most impact on the output.

[0062] After selecting 8 or more features, a convergence toward the most important features is observed. The K features having the most counts for K are going to be used as sample features to test the accuracy of the different learning algorithms.

[0063] FIG. 4 is a schematic 400 that shows exemplary results for a plurality of machine learning functions as a function of the training length according to one example. Schematic 400 represents the accuracy of each algorithm shown in Table 2 depending on the number of features selected from the data set. al in FIG. 4 represents the accuracy using the EMPC+ average load forecast algorithm for mode selection. The exact same features are extracted for each house, based on the features getting the most counts for each k in table 5. The accuracy increases with the number of feature. Starting at around 8 features, the marginal accuracy gain starts to decrease and converge.

[0064] FIG. 5 is a schematic 500 that shows exemplary results for the plurality of machine learning functions as a function of the number of features in a training set. Schematic 500 represents the accuracy of each algorithm function of the length of the training set. The length goes from one month, 30 days, to 11 months, 330 days. The testing month is August. The months are added to the training set from July to September. In this case, all the algorithms are tested with the 14 features scenario. Two observations can be made: the SVM and smaller NN perform better with less data, and start converging with 3 months of data.

The NN using 10 neurons in one or two hidden layers, and the NN using 20 neurons in one hidden layer perform better with more data. Those networks do not seem to converge with the amount of data proposed. The NN using 20 neurons in one layer achieves the highest accuracy. In one implementation, the NN using 20 neurons is used by the controller 104. [0065] FIG. 6 is a schematic 600 that shows the accuracy for each mode according to one example. The precision for each mode selection is compared with the EMPC-based mode algorithm. The ML process is implemented using the 14 best features scenario over 11 months of data. Each value represents the mean of the result obtained over the whole dataset (149 houses). The overall precision is higher in the machine learning case. For the mode 0, the performance is equivalent for both algorithms. However, the ML-based algorithm significantly increased the accuracy for the mode 1, 2 and 3. Mode 4 is never selected, as in self-consumption case the backfeed is not valorized.

[0066] The test month is August, and the training set contains the remaining months of the year. The case with MILP + perfect forecast is assumed to achieve the most savings (100%). In one implementation, the NN with 20 neurons in one hidden layer using 14 features is used. The training length is 11 months.

Table 5: K best feature selected

[0067] IG. 7 is a schematic 700 that shows the energy cost for machine learning mode based control and for economic model predictive control (EMPC) mode based control. Schematic 700 shows the results of the simulation comparing the EMPC-based mode control (702) with the ML-based mode control algorithm (704). The ML-based algorithm yields

better performance. The EMPC algorithm (702) achieved on average 73% of the potential maximum savings achievable while the ML approach (704) yielded 82%.

[0068] The features of the present disclosure provide a multitude of improvements in the technical field of battery management. In particular, the controller identifies an operation mode for the ESD for the next predetermined period using a machine learning approach which reduce the reliance on PV and load forecasters. The methodology described herein achieves high accuracy in mode prediction while minimizing the computation requirement by eliminating an optimization step for each operating period. The methodology described herein could not be implemented by a human due to the sheer complexity of data, gathering and calculating and includes a variety of novel features and elements that result is

significantly more than an abstract idea. The ESD may be standardized. Further, the safety and efficiency of the ESD is improved as the operation conditions of each mode can be predefined. Thus, the implementations described herein improve the functionality of the ESD. Thus, the system and associated methodology described herein amount to significantly more than an abstract idea based on the improvements and advantages described herein.

[0069] In one implementation, the functions and processes of the controller 104 may be implemented by a computer 826. Next, a hardware description of the computer 826 according to exemplary embodiments is described with reference to FIG. 8. In FIG. 8, the computer 826 includes a CPET 800 which performs the processes described herein. The process data and instructions may be stored in memory 802. These processes and instructions may also be stored on a storage medium disk 804 such as a hard drive (HDD) or portable storage medium or may be stored remotely. Further, the claimed advancements are not limited by the form of the computer-readable media on which the instructions of the inventive process are stored. For example, the instructions may be stored on CDs, DVDs, in FLASH memory, RAM, ROM, PROM, EPROM, EEPROM, hard disk or any other information processing device with which the computer 826 communicates, such as a server or computer.

[0070] Further, the claimed advancements may be provided as a utility application, background daemon, or component of an operating system, or combination thereof, executing in conjunction with CPET 800 and an operating system such as Microsoft® Windows®, UNIX®, Oracle® Solaris, LINUX®, Apple macOS® and other systems known to those skilled in the art.

[0071] In order to achieve the computer 826, the hardware elements may be realized by various circuitry elements, known to those skilled in the art. For example, CPU 800 may be a Xenon® or Core® processor from Intel Corporation of America or an Opteron® processor from AMD of America, or may be other processor types that would be recognized by one of ordinary skill in the art. Alternatively, the CPU 800 may be implemented on an FPGA,

ASIC, PLD or using discrete logic circuits, as one of ordinary skill in the art would recognize. Further, CPU 800 may be implemented as multiple processors cooperatively working in parallel to perform the instructions of the inventive processes described above.

[0072] The computer 826 in FIG. 8 also includes a network controller 806, such as an Intel Ethernet PRO network interface card from Intel Corporation of America, for interfacing with network 824. As can be appreciated, the network 824 can be a public network, such as the Internet, or a private network such as LAN or WAN network, or any combination thereof and can also include PSTN or ISDN sub-networks. The network 824 can also be wired, such as an Ethernet network, or can be wireless such as a cellular network including EDGE, 3G and 4G wireless cellular systems. The wireless network can also be WiFi®, Bluetooth®, or any other wireless form of communication that is known.

[0073] The computer 826 further includes a display controller 808, such as a NVIDIA® GeForce® GTX or Quadro® graphics adaptor from NVIDIA Corporation of America for interfacing with display 810, such as a Hewlett Packard® HPL2445w LCD monitor. A general purpose I/O interface 812 interfaces with a keyboard and/or mouse 814 as well as an optional touch screen panel 816 on or separate from display 810. General purpose I/O interface also connects to a variety of peripherals 818 including printers and scanners, such as an OfficeJet® or DeskJet® from Hewlett Packard®.

[0074] The general purpose storage controller 820 connects the storage medium disk 804 with communication bus 822, which may be an ISA, EISA, VESA, PCI, or similar, for interconnecting all of the components of the computer 826. A description of the general features and functionality of the display 810, keyboard and/or mouse 814, as well as the display controller 808, storage controller 820, network controller 806, and general purpose I/O interface 812 is omitted herein for brevity as these features are known.

[0075] Obviously, numerous modifications and variations are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims, the invention may be practiced otherwise than as specifically described herein.

[0076] Thus, the foregoing discussion discloses and describes merely exemplary

embodiments of the present invention. As will be understood by those skilled in the art, the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. Accordingly, the disclosure of the present invention is intended to be illustrative, but not limiting of the scope of the invention, as well as other claims. The disclosure, including any readily discernible variants of the teachings herein, defines, in part, the scope of the foregoing claim terminology such that no inventive subject matter is dedicated to the public.

[0077] The above disclosure also encompasses the embodiments listed below.

[0078] (1) An energy storage device, including a controller configured to identify a predetermined number of features based on forecasted data and data obtained from the energy storage device; identify an operating mode from a plurality of operating modes for a predetermined operating period using a machine learning model; and operate the energy storage device in the operating mode for the predetermined operating period.

[0079] (2) The device of feature (1), in which the plurality of operating modes include a charge mode from energy generated from the at least one renewable energy source, another charge mode from energy generated from an energy grid, a discharge mode at a maximum power, another discharge mode following the load profile, and an idle mode.

[0080] (3) The device of feature (1) or (2), in which the controller is further configured to update the machine learning model when training data associated with the machine learning model is not updated within a predetermined period.

[0081] (4) The device of any of feature (1) to (3), in which updating the machine learning model includes the steps of: determine an operating schedule for the energy storage device based on at least an initial optimization parameter, the initial optimization parameter being a function of at least a load profile and a power production profile of at least one renewable energy source associated with the energy storage device and the operating schedule being established for a predetermined period; determine an updated optimization parameter based on at least the operating schedule; identify an optimal operating mode from a plurality of operating modes for a predetermined operating period based on the updated optimization parameter and the operating schedule; and update the machine learning model based on the optimal operating mode for each predetermined operating period.

[0082] (5) The device of any of feature (1) to (4), in which the controller is further configured to store data from a previous predetermined operating period; obtain forecasted data including at least a temperature profile; obtain data from the energy storage device including at least a state of charge associated with the energy storage device; identify an optimal operating mode from the plurality of operating modes for the predetermined operating period; operate the energy storage device in the optimal operating mode for the predetermined operating period; repeat the steps of store, operate, and identify until the number of predetermined operating periods exceeds a predefined threshold; and store the optimal operating mode for each of the predetermined operating periods.

[0083] (6) The device of any of features (1) to (5), in which in response to determining that the stored predetermined operating periods exceeds the predefined threshold, the controller is further configured to identify a second optimal mode for each of the stored predetermined operating mode using at least a perfect forecaster; create a first vector in a first register of the controller containing identified features; create a second vector in a second register of the controller containing the second optimal modes associated with the first vector; and train the machine learning model using the first vector and the second vector.

[0084] (7) The device of any of features (1) to (6), in which the controller is further configured to forecast the load profile for the predetermined period using a neural network forecaster.

[0085] (8) The device of any of features (1) to (7), in which the controller is further configured to forecast a power-production profile of the at least one renewable energy source associated with the energy storage device for the predetermined period.

[0086] (9) The device of any of the features (1) to (7), in which the at least one renewable energy source is a photovoltaic system.

[0087] (10) The device of any of the features (1) to (9), in which the predetermined operating period is one hour.

[0088] (11) A method for controlling an energy storage device, the method including identifying, using processing circuitry, a predetermined number of features based on forecasted data and data obtained from the energy storage device; identifying, using the processing circuitry, an operating mode from a plurality of operating modes for a predetermined operating period using a machine learning model; and operating the energy storage device in the operating mode for the predetermined operating period.

[0089] (12) The method of feature (11), in which the plurality of operating modes include a charge mode from energy generated from the at least one renewable energy source, another charge mode from energy generated from an energy grid, a discharge mode at a maximum power, another discharge mode following the load profile, and an idle mode.

[0090] (13) The method of feature (11) or (12), further including updating the machine learning model when training data associated with the machine learning models is not updated within a predetermined period.

[0091] (14) The method of any of features (11) to (13), in which updating the machine learning model includes the steps of: determining an operating schedule for the energy storage device based on at least an initial optimization parameter, the initial optimization parameter being a function of at least a load profile and a power production profile of at least one renewable energy source associated with the energy storage device and the operating schedule being established for a predetermined period; determining an updated optimization parameter based on at least the operating schedule; identifying an optimal operating mode from a plurality of operating modes for a predetermined operating period based on the updated optimization parameter and the operating schedule; and updating the machine learning model based on the optimal operating mode for each predetermined operating period.

[0092] (15) The method of any of features (11) to (14), further including: storing data from a previous predetermined operating period; obtaining forecasted data for at least a temperature profile; obtaining data from the energy storage device including at least a state of charge associated with the energy storage device; identifying an optimal operating mode from the plurality of operating modes for the predetermined operating period; operating the energy storage device in the operating mode for the predetermined operating period; repeating the steps of storing, operating, and identifying until the number of predetermined operating periods exceeds a predefined threshold; and storing the optimal operating mode for each of the predetermined operating periods.

[0093] (16) The method of any of features (11) to (15), further including: identifying a second optimal mode for each of the stored predetermined operating mode using at least a perfect forecaster in response to determining that the stored predetermined operating periods exceeds the predefined threshold; creating a first vector in a first register containing identified features; creating a second vector in a second register containing the second optimal modes associated with the first vector; and training the machine learning model using the first vector and the second vector.

[0094] (17) The method of any of features (11) to (15), further including forecasting the load profile for the predetermined period using a neural network forecaster.

[0095] (18) The method of any of features (11) to (17), in which the predetermined operating period is one hour.

[0096] (19) A system including an energy storage device configured to operate in an operating mode indicated by a control signal received from a controller and the controller.

The controller is configured to identify a predetermined number of features based on forecasted data and data obtained from the energy storage device; identify an operating mode from a plurality of operating modes for a predetermined operating period using a machine learning model; and operate the energy storage device in the operating mode for the predetermined operating period.

[0097] (20) The system of feature (19), in which the plurality of operating modes include a charge mode from energy generated from the at least one renewable energy source, another charge mode from energy generated from an energy grid, a discharge mode at a maximum power, another discharge mode following the load profile, and an idle mode.

[0098] (21) The system of feature (19) or (20), in which the controller is further configured to update the machine learning model when training data associated with the machine learning model is not updated within a predetermined period.

[0099] (22) The system of any of feature (19) to (21), in which updating the machine learning model includes the steps of: determine an operating schedule for the energy storage device based on at least an initial optimization parameter, the initial optimization parameter being a function of at least a load profile and a power production profile of at least one renewable energy source associated with the energy storage device and the operating schedule being established for a predetermined period; determine an updated optimization parameter based on at least the operating schedule; identify an optimal operating mode from a plurality of operating modes for a predetermined operating period based on the updated optimization parameter and the operating schedule; and update the machine learning model based on the optimal operating mode for each predetermined operating period.

[00100] (23) The system of any of feature (19) to (22), in which the controller is further configured to store data from a previous predetermined operating period; obtain forecasted data including at least a temperature profile; obtain data from the energy storage device including at least a state of charge associated with the energy storage device; identify an optimal operating mode from the plurality of operating modes for the predetermined operating period; operate the energy storage device in the optimal operating mode for the predetermined operating period; repeat the steps of store, operate, and identify until the number of predetermined operating periods exceeds a predefined threshold; and store the optimal operating mode for each of the predetermined operating periods. [00101] (24) The system of any of features (19) to (23), in which in response to determining that the stored predetermined operating periods exceeds the predefined threshold, the controller is further configured to identify a second optimal mode for each of the stored predetermined operating mode using at least a perfect forecaster; create a first vector in a first register of the controller containing identified features; create a second vector in a second register of the controller containing the second optimal modes associated with the first vector; and train the machine learning model using the first vector and the second vector.

[00102] (25) The system of any of features (19) to (24), in which the controller is further configured to forecast the load profile for the predetermined period using a neural network forecaster.

[00103] (26) The system of any of features (19) to (25), in which the controller is further configured to forecast a power-production profile of the at least one renewable energy source associated with the energy storage device for the predetermined period.

[00104] (27) The system of any of the features (19) to (26), in which the at least one renewable energy source is a photovoltaic system.

[00105] (28) The system of any of the features (19) to (27), in which the predetermined operating period is one hour.

Claims

1. An energy storage device, comprising:

a controller configured to

identify a predetermined number of features based on forecasted data and data obtained from the energy storage device;

identify an operating mode from a plurality of operating modes for a predetermined operating period using a machine learning model; and

operate the energy storage device in the operating mode for the predetermined operating period.

2. The device of claim 1, wherein the plurality of operating modes include a charge mode from energy generated from the at least one renewable energy source, another charge mode from energy generated from an energy grid, a discharge mode at a maximum power, another discharge mode following the load profile, and an idle mode.

3. The device of claim 1, wherein the controller is further configured to update the machine learning model when training data associated with the machine learning model is not updated within a predetermined period.

4. The device of claim 3, wherein updating the machine learning model includes the steps of:

determine an operating schedule for the energy storage device based on at least an initial optimization parameter, the initial optimization parameter being a function of at least a load profile and a power production profile of at least one renewable energy source associated with the energy storage device and the operating schedule being established for a predetermined period;

determine an updated optimization parameter based on at least the operating schedule; identify an optimal operating mode from a plurality of operating modes for a predetermined operating period based on the updated optimization parameter and the operating schedule; and update the machine learning model based on the optimal operating mode for each predetermined operating period.

5. The device of claim 1, wherein the controller is further configured to:

store data from a previous predetermined operating period;

obtain forecasted data including at least a temperature profile;

obtain data from the energy storage device including at least a state of charge associated with the energy storage device;

identify an optimal operating mode from the plurality of operating modes for the predetermined operating period;

operate the energy storage device in the optimal operating mode for the predetermined operating period;

repeat the steps of store, operate, and identify until the number of predetermined operating periods exceeds a predefined threshold; and

store the optimal operating mode for each of the predetermined operating periods.

6. The device of claim 5, wherein in response to determining that the stored

predetermined operating periods exceeds the predefined threshold, the controller is further configured to:

identify a second optimal mode for each of the stored predetermined operating mode using at least a perfect forecaster;

create a first vector in a first register of the controller containing identified features;

create a second vector in a second register of the controller containing the second optimal modes associated with the first vector; and

train the machine learning model using the first vector and the second vector.

7. The device of claim 5, wherein the controller is further configured to forecast the load profile for the predetermined period using a neural network forecaster.

8. The device of claim 5, wherein the controller is further configured to forecast a

power-production profile of the at least one renewable energy source associated with the energy storage device for the predetermined period.

9. The device of claim 8, wherein the at least one renewable energy source is a

photovoltaic system.

10. The device of claim 1, wherein the predetermined operating period is one hour.

11. A method for controlling an energy storage device, the method comprising: identifying, using processing circuitry, a predetermined number of features based on forecasted data and data obtained from the energy storage device;

identifying, using the processing circuitry, an operating mode from a plurality of operating modes for a predetermined operating period using a machine learning model; and

operating the energy storage device in the operating mode for the predetermined operating period.

12. The method of claim 11, wherein the plurality of operating modes include a charge mode from energy generated from the at least one renewable energy source, another charge mode from energy generated from an energy grid, a discharge mode at a maximum power, another discharge mode following the load profile, and an idle mode.

13. The method of claim 11, further comprising updating the machine learning model when training data associated with the machine learning models is not updated within a predetermined period.

14. The method of claim 13, wherein updating the machine learning model includes the steps of:

determining an operating schedule for the energy storage device based on at least an initial optimization parameter, the initial optimization parameter being a function of at least a load profile and a power production profile of at least one renewable energy source associated with the energy storage device and the operating schedule being established for a predetermined period;

determining an updated optimization parameter based on at least the operating schedule; identifying an optimal operating mode from a plurality of operating modes for a predetermined operating period based on the updated optimization parameter and the operating schedule; and updating the machine learning model based on the optimal operating mode for each predetermined operating period.

15. The method of claim 11, further comprising:

storing data from a previous predetermined operating period;

obtaining forecasted data for at least a temperature profile;

obtaining data from the energy storage device including at least a state of charge associated with the energy storage device;

identifying an optimal operating mode from the plurality of operating modes for the predetermined operating period;

operating the energy storage device in the operating mode for the predetermined operating period; repeating the steps of storing, operating, and identifying until the number of predetermined operating periods exceeds a predefined threshold; and

storing the optimal operating mode for each of the predetermined operating periods.

16. The method of claim 15, further comprising:

identifying a second optimal mode for each of the stored predetermined operating mode using at least a perfect forecaster in response to determining that the stored predetermined operating periods exceeds the predefined threshold;

creating a first vector in a first register containing identified features; creating a second vector in a second register containing the second optimal modes associated with the first vector; and

training the machine learning model using the first vector and the second vector.

17. The method of claim 15, further comprising forecasting the load profile for the

predetermined period using a neural network forecaster.

18. The method of claim 11, wherein the predetermined operating period is one hour.

19. A system, comprising:

an energy storage device configured to operate in an operating mode indicated by a control signal received from a controller; and

the controller configured to identify a predetermined number of features based on forecasted data and data obtained from the energy storage device;