Method and System for ESS AI Dynamic Operation Combining MILP Optimization and Kalman-Filter-Based Rolling Horizon under Time-of-Use Tariff

Abstract

The present invention discloses an ESS AI Dynamic operation method characterized by comprising: (a) a step of predicting PV generation amount and load demand; (b) a step of calculating a charge / discharge schedule for a 24-hour rolling horizon through Mixed Integer Linear Programming (MILP) by reflecting the unit price of the Korea Electric Power Corporation rate table in a cost function; (c) a step of smoothing real-time PV generation amount and load values ​​measured at a predetermined period using a Kalman filter; (d) a step of real-time correction of the charge / discharge schedule to the rolling horizon when the deviation between the real-time PV generation amount and load values ​​and the charge / discharge schedule exceeds a predetermined threshold; and (e) a step of transmitting the corrected command value to a power conversion device (PCS).

Description

Technology Field

[0001] The present invention relates to an automatic operation method and system for an energy storage system (hereinafter referred to as 'ESS'), and more specifically, to an ESS AI Dynamic operation method and system that simultaneously achieves PV self-consumption rate and peak reduction rate while minimizing electricity costs by predicting PV (Photovoltaic) generation and load demand in 24-hour units under a seasonal and hourly electricity rate system, calculating a daily optimal charge / discharge schedule using MILP (Mixed-Integer Linear Programming), removing noise from real-time measurements using a Kalman Filter, and correcting every hour using the Rolling Horizon technique. Background Technology

[0002] Conventional ESSs generally perform charging and discharging only according to the fixed rules of the selected mode when the user manually selects and sets the operating mode. The following six representative existing operating modes are applied to ESSs previously supplied by the applicant and ESSs in the same industry.

[0003] (i) Self Use with PV Power Mode ― Priority Load > Battery > Grid, PV power generation is prioritized for self-consumption, and surplus power is transmitted to the grid after charging the battery.

[0004] (ii) Self Use without PV Power mode ― Priority load > Battery. When there is no or insufficient PV, the battery supplies power to the load first, and when the battery is insufficient, the grid supplies power as a backup.

[0005] (iii) TOU Balance Mode ― When the user directly inputs the time period (peak / flat / valley) and the time period for sufficient PV generation, the peak time period is self-consumption mode, the flat time period is PV priority load supply, and the valley time period is charged until the battery is fully charged.

[0006] (iv) TOU Eco Mode ― In areas with a large difference between peak and valley prices, the battery is discharged at maximum output during peak hours and transmitted to the grid (sold), consumed during flat hours, and fully charged at a low price from the grid during valley hours.

[0007] (v) Backup only mode ― Priority load > Battery, the battery is always kept fully charged and emergency power is supplied from the PV battery only in the event of a power outage.

[0008] (vi) Grid Priority mode ― Priority load > Grid, PV generation power is transmitted to the grid first, and the user can manually send transmission requests during peak hours and the battery SOC endpoint is also manually set by the user.

[0009] All of the existing modes mentioned above have limitations in that (a) the user must manually set the mode and mode-specific parameters (time zone, SOC end point, sales price threshold, etc.) in advance, (b) since they operate in a fixed manner according to a rule set once, they cannot reflect daily fluctuations in PV generation (clouds, rain, etc.) or sudden increases or decreases in load, and (c) they cannot explicitly reflect the asymmetry of unit prices by time zone, season, and holiday in the KEPCO rate table as a cost function.

[0010] In particular, as KEPCO’s Industrial (Type B) High Voltage A Option II rate system is set to be implemented on April 16, 2026, and unit prices fluctuate by up to about two times depending on the time of day, season, and holidays, the savings effect has become limited with fixed rule methods such as the existing modes mentioned above.

[0011] Patent Document 1 presented a charging and discharging control method based on Model Predictive Control (MPC) for a Battery Energy Storage System (BESS), but this was limited to short-period control based on a numerical model in minutes, and did not cover daily optimization that integrates time-based price differences of a time-based tariff system into an explicit cost function or combination with a Kalman filter that removes measurement noise, nor did it present an artificial intelligence operation method that automatically operates based on input variables while only requiring the user to select a mode.

[0012] Other related prior art includes Patent Document 2 (Power system frequency control using SoC of multiple ESSs), which has a different purpose from the present invention and does not address time-of-use rate cost functions; Patent Document 3 (Operation method for large-capacity BESS peak shaving), which presents next-day peak discharge scheduling based on SMP (System Marginal Price) and real-time re-execution based on EMPC, but does not directly reflect the asymmetric unit prices for season, holidays, and weekends in the KEPCO industrial time-of-use rate table in the cost function. In addition, Patent Document 4 (Time-of-Use based ESS operation), Patent Document 5 (Demand Charge), and Patent Document 6 (Rolling Horizon Optimizer) are designed for US-style TOU rates and do not reflect the asymmetric structure of time-of-use, seasonal, weekend, and holiday unit prices in the Korean industrial (Type B) rate table, and do not present any algorithms for processing measurement noise through Kalman filters.

[0013] Considering the above points, a new algorithm is required that combines (i) proactive and planned 24-hour daily optimization, (ii) real-time and responsive minute-by-minute correction, (iii) stably without measurement noise, and (iv) automatically minimizes electricity costs without separate manual settings, provided the user simply inputs the input variables and selects "AI Dynamic Mode". Prior art literature

[0014] Korean Registered Patent No. 10-1786904 (Registered Oct. 11, 2017) Korean Registered Patent No. 10-1697348 (Registered Jan. 11, 2017) Korean Published Patent No. 10-2016-0017682 (Published Feb. 17, 2016) US Registered Patent US 11,476,693 (Registered Oct. 18, 2022) US Registered Patent US 10,673,241 (Registered Jun. 2, 2020) US Registered Patent US 10,673,242 (Registered Jun. 2, 2020) The problem to be solved

[0015] The present invention was devised to solve the above-mentioned problems and aims to solve the following problems.

[0016] First, by directly reflecting the differences in unit prices by time of day, season, and holiday—specifically the seasonal and hourly rate systems representative of Korea—into the cost function, the savings rate compared to a fixed schedule is increased.

[0017] Second, it absorbs daily fluctuations in PV generation and load variations in real time while maintaining the stability of hourly decision-making.

[0018] Third, noise is suppressed using a Kalman filter so that the algorithm does not overreact to noise from the measurement sensor.

[0019] Fourth, it provides a multi-stage safety structure that can safely fallback even in abnormal situations such as MILP solver failure, communication loss, or SoC limit violation.

[0020] Fifth, the user does not have to perform separate manual settings (time zone input, SOC endpoint setting, electricity price threshold setting, etc.), but simply inputs the site's input variables (PV capacity, load pattern identifier, tariff code, ESS model and number of units, etc.) into the server and selects "AI Dynamic Mode," and the algorithm of the present invention automatically calculates and applies an optimal charge / discharge schedule through artificial intelligence, thereby reducing the operational burden on the user. means of solving the problem

[0021] To achieve the above objectives, the present invention provides an ESS AI Dynamic operation method characterized by comprising: (a) a step of predicting PV generation amount and load demand; (b) a step of calculating a 24-hour rolling horizon charging and discharging schedule through Mixed Integer Linear Programming (MILP) by reflecting the unit price of the Korea Electric Power Corporation rate table in a cost function; (c) a step of smoothing real-time PV generation amount and load values ​​measured at a predetermined period using a Kalman filter; (d) a step of real-time correction of the charging and discharging schedule to the rolling horizon when the deviation between the real-time PV generation amount and load values ​​and the charging and discharging schedule exceeds a predetermined threshold; and (e) a step of transmitting the corrected command value to a power conversion device (PCS).

[0022] Step (a) above obtains a 24-hour PV power generation time series from the Korea Meteorological Administration short-term forecast data, predicts the PV power generation and load demand through five physical model steps of external solar radiation, panel incidence angle correction, attenuation due to cloud cover, PV cell temperature correction, and inverter efficiency application, and the load demand time series can be predicted using a machine learning model.

[0023] The above step (b) may further include the step of calculating the charge / discharge schedule by reflecting unit prices by time period, season, and holiday—including discount unit prices for spring, autumn, Saturday, Sunday, and public holidays in the Korea Electric Power Corporation rate table, the reduction of Saturday maximum load to an intermediate load, and the application of low load before Sunday and public holidays—into a cost function, and performing a heuristic fallback in which, if the solver of the MILP fails to produce a solution within a certain period of time, time periods when the unit price by time is lower than the daily average are assigned as the charging period and time periods when it is higher are assigned as the discharging period.

[0024] Step (c) above can smooth real-time PV·load values ​​measured at 1-second intervals using a Kalman filter that uses Joseph form covariance update.

[0025] Step (d) above can correct the charge / discharge schedule by re-executing the MILP if the deviation between the actual energy integral value of the previous hour and the charge / discharge schedule exceeds a predetermined threshold at every hour.

[0026] Step (e) above can transmit the corrected command value to the Power Conversion System (PCS) via the Modbus TCP protocol after passing it through the safety gates of State of Charge (SoC), State of Health (SoH), and battery temperature.

[0027] If the PCS communication in step (e) above is lost for more than a certain period of time, a Safe Idle command (p_ref=0, mode=0) can be automatically issued to switch the ESS to a safe stop state.

[0028] When a user inputs site-specific input variables—including PV capacity, load pattern identifier, KEPCO rate plan code, ESS model, and number of parallel units—once into the operating server and selects "AI Dynamic Mode," the operating server can automatically repeat the prediction of step (a), the 24-hour MILP optimization of step (b), the rolling horizon correction of step (d), and the PCS command transmission of step (e) without any additional manual settings by the user, thereby minimizing electricity costs.

[0029] According to another aspect of the present invention, an ESS AI Dynamic operating system is provided, comprising: a prediction unit that predicts PV generation amount and load demand in a system for automatically operating the charging and discharging of an ESS in a seasonal and hourly rate system environment; an optimization unit that calculates a charging and discharging schedule for a 24-hour rolling horizon through Mixed Integer Linear Programming (MILP) by reflecting the unit price of the Korea Electric Power Corporation rate table in a cost function; a correction unit that smooths real-time PV generation amount and load values ​​measured at a fixed period using a Kalman filter and corrects the charging and discharging schedule in real-time to a rolling horizon when the deviation between the real-time PV generation amount and load values ​​and the charging and discharging schedule exceeds a fixed threshold; an adapter unit capable of communicating with a PCS via Modbus TCP; and a safety supervision unit that monitors State of Charge (SoC), State of Health (SoH), and battery temperature limits to block abnormal commands.

[0030] Each of the above-mentioned units is executed as an independent process, configured so that the safety supervision unit alone can issue a Safe Idle command even if a failure occurs in any of the units, and the same algorithm is commonly applied to commercial and residential ESSs with variable PCS and battery capacities, and when multiple units are connected in parallel, the site rating is converted and operated with the same algorithm.

[0031] According to another aspect of the present invention, a computer-readable recording medium is provided that has a program capable of executing the following: a procedure for predicting PV generation amount and load demand; a procedure for calculating a 24-hour rolling horizon charging and discharging schedule through Mixed Integer Linear Programming (MILP) by reflecting the unit price of the Korea Electric Power Corporation rate table in a cost function; a procedure for smoothing real-time PV generation amount and load values ​​measured at a fixed period using a Kalman filter; a procedure for correcting the charging and discharging schedule in real-time to the rolling horizon when the deviation between the PV generation amount and load real-time values ​​and the charging and discharging schedule exceeds a fixed threshold; and a procedure for transmitting the corrected command value to a power conversion device (PCS). Effects of the invention

[0032] The ESS AI Dynamic operation method and system according to the present invention have the following effects.

[0033] First, through a cost function that reflects time zones, seasons, and holidays and 24-hour MILP optimization, annual electricity costs can be reduced by an average of more than 18% compared to a fixed schedule.

[0034] Second, by suppressing measurement noise with a Kalman filter and correcting with a Rolling Horizon, it is possible to simultaneously ensure the stability of daily planning and real-time adaptability.

[0035] Third, system availability of over 99.5% can be secured through a four-stage fallback of MILP failure → heuristic → TOU fix → Safe Idle.

[0036] Fourth, the same algorithm can be applied to both 60kW commercial ESS and 10kW residential ESS, and can be converted and supported when 2 to 10 units are installed and operated in parallel.

[0037] Fifth, the user only needs to input site-specific input variables (PV capacity, load pattern identifier, tariff code, ESS model and number of units, etc.) into the server once and select "AI Dynamic Mode," after which all subsequent charge and discharge decisions can be automatically performed by the algorithm of the present invention. Unlike the six conventional modes (Self Use with PV / Self Use without PV / TOU Balance / TOU Eco / Back up only / Grid Priority) where the user had to manually set the time zone, SOC endpoint, and electricity price threshold, the present invention does not require any such manual setting at all, and as a result, the user's operational burden is reduced to 0% while the savings can be automatically maximized. Brief explanation of the drawing

[0038] The following drawings attached to this specification illustrate preferred embodiments of the present invention and serve to further enhance understanding of the technical concept of the present invention together with the detailed description of the invention provided below; therefore, the present invention should not be interpreted as being limited only to the matters described in such drawings. FIG. 1 is a four-layer block diagram of an ESS AI Dynamic operating system according to a preferred embodiment of the present invention. Figure 2 is a 5-step flowchart of the Layer 0 PV prediction physical model. Figure 3 is a Layer 1 MILP daily optimization flowchart. Figure 4 is a flowchart of the prediction-update cycle of a Layer 2 Kalman filter. Figure 5 is a Layer 2 Rolling Horizon trigger condition and re-optimization flowchart. Figure 6 is a state diagram showing the response to communication loss of a Layer 3 PCS Modbus adapter. Figure 7 is an example of a 24-hour simulation comparison graph (rule-based vs. the present invention). FIG. 8 is a system configuration diagram (DC coupled method) of a Hybrid ESS applied to the present invention. Figure 9 is a communication connection diagram between the GLS cloud server and the ESS. Specific details for implementing the invention

[0039] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings.

[0040] FIG. 1 is a four-layer (Layer 0~3) block diagram of an ESS AI Dynamic operating system according to a preferred embodiment of the present invention.

[0041] Referring to FIG. 1, an ESS AI Dynamic operating system according to a preferred embodiment of the present invention may provide an operating server (200) comprising a prediction unit (210), an optimization unit (220), a correction unit (230), an adapter unit (240), and a safety supervision unit (250). Each unit is executed separately as a separate process or container and can communicate through a message queue (MQ).

[0042] The ESS (120) to which the present invention is applied can be configured with a DC-coupled hybrid structure as shown in FIG. 8. DC power generated from the PV module (112) is directly input to the DC bus of the Hybrid PCS (122), and the same PCS simultaneously performs charging and discharging of the battery (124), transmission and reception to the grid (140), and supply to the load (130) through bidirectional conversion. The PMS (126) acts as an integrated energy management, user convenience, and communication gateway above the PCS (122). This integrated DC-coupled configuration reduces the conversion step by one compared to the conventional AC-coupled structure (where the solar inverter and ESS PCS are separated), thereby reducing conversion losses and reducing installation, transportation, and maintenance costs.

[0043] Layer 0 (prediction layer) receives hourly cloud cover, solar radiation, and temperature data from the Korea Meteorological Administration's short-term forecast, predicts a 24-hour time series of PV power generation using a 5-stage physical model, and predicts a 24-hour time series of load demand using a LightGBM or LSTM model. That is, the prediction unit (210) constituting Layer 0 calls the Korea Meteorological Administration's short-term forecast API, for example, at 23:30 every day, to obtain the time series of cloud cover (0~10), solar radiation (W / m²), temperature (°C), and wind speed (m / s) for the next 24 hours. As shown in FIG. 2, the 5-stage PV prediction physical model consists of 5 stages: (i) outdoor solar radiation → (ii) panel incidence angle correction → (iii) attenuation due to cloud cover → (iv) cell temperature correction → (v) inverter efficiency application. The coefficients for each stage can be operated as the site default value + 30-day correction value after installation.

[0044] Layer 1 (MILP optimization layer) can be configured to solve a MILP problem that minimizes, for example, the following cost function over a 24-hour horizon at midnight every day.

[0045] minimize Σ[ p_grid(t) × tariff(t) ] + λ× peak_kW

[0046] subject to:

[0047] p_pv(t) - p_load(t) - p_ess(t) - p_grid(t) = 0 (Power Balance)

[0048] SoC(t+1) = SoC(t) + η_c x p_chg(t) x Δt - p_dis(t) Х Δt / η_d

[0049] SoC_min ≤ SoC(t) ≤ SoC_max

[0050] -P_pcs ≤ p_ess(t) ≤ +P_pcs

[0051] p_chg(t) x p_dis(t) = 0 (Simultaneous charging and discharging prohibited, expressed as a binary variable)

[0052] Here, tariff(t) is a unit price function reflecting time of day, season, and holidays based on the KEPCO Industrial (Type B) High Voltage A Select II Rate Table (Korea’s representative electricity rate system in 2026), and a 50% discount is automatically applied from 11 to 14:00 on Saturdays, Sundays, and public holidays in spring and autumn, a downgrade from peak load to intermediate load is automatically performed on Saturdays, and off-peak load is automatically applied for all time slots on Sundays and public holidays.

[0053] The optimization unit (220) constituting Layer 1 solves MILP for 24 time variables (p_chg, p_dis, soc, p_grid) at 00:00 every day, for example. As shown in FIG. 3, the solver uses CBC (COIN-OR Branch-and-Cut) as the first priority and GLPK (GNU Linear Programming Kit) as the second priority, and if it does not obtain a solution within 30 seconds, it falls back to the following heuristic.

[0054] Heuristic: Designate the bottom 8 hours as the charging period and the top 6 hours as the discharging period among the time periods when tariff(t) is lower than the daily average. Normalize the sum of charges and discharges to match the available capacity of the site.

[0055] Layer 2 (real-time correction layer) consists of the following two parts.

[0056] (a) Kalman filter: PV·load measurements received at 1-second intervals are smoothed using a 2-state (p_pv, p_load) Kalman filter. The state transition is assumed to be a random walk (F=I), and process noise Q and measurement noise R are learned on a site-by-site basis. Covariance updates are performed using the Joseph form, which guarantees numerical stability.

[0057] (b) Rolling Horizon: At every hour on the hour, calculate the deviation between the actual value from the previous hour and the Layer 1 pre-schedule. If the deviation exceeds ±30%, re-run the shortened MILP for the next 4 hours of the horizon to update the command value. Otherwise, maintain the pre-schedule.

[0058] The correction unit (230) constituting Layer 2 operates, for example, a 1-second period fast tick and a regular hourly tick simultaneously.

[0059] Figure 4 illustrates the prediction-update cycle of the Kalman filter (Joseph form covariance update).

[0060] Fast tick: Input the measurements (p_pv_meas, p_load_meas) into the Kalman filter.

[0061] Joseph form update: P_post = (I - K x H) x P_prior x (I - K x H)^T + K x R x K^T

[0062] The above format prevents asymmetric and non-head portions caused by arithmetic errors and ensures that P maintains head symmetry even during long-term operation.

[0063] Figure 5 illustrates the Rolling Horizon trigger condition and the re-optimization process.

[0064] Hourly tick: Calculates the difference between the actual integral value of the previous hour and the Layer 1 pre-schedule.

[0065] deviation = | (E_actual - E_planned) / E_planned |

[0066] If deviation > 0.30, run a shortened MILP with the next 4 hours as the horizon to partially update the prior schedule.

[0067] Layer 3 (PCS adapter layer) transmits command values ​​(p_ref, q_ref, mode) to the PCS via Modbus TCP and retrieves measurement values ​​(p_meas, soc, soh, alarm). If communication is lost for more than 5 seconds, a Safe Idle (p_ref=0) command is issued immediately. The adapter unit (240) constituting Layer 3 uses standardized Modbus register mapping (p_ref=0x1000, q_ref=0x1002, mode=0x1004, p_meas=0x2000, soc=0x2004, alarm=0x2010, soh=0x2014). As shown in FIG. 6, communication is polled at a 1-second interval, and when 5 seconds have elapsed from the last response time, the safety supervisor unit (250) issues a mode=0 (Safe Idle), p_ref=0 command.

[0068] The operating server (200) provided according to the present invention operates, for example, on a GLS monitoring cloud server (310, https: / / k-gls.com) and is connected to the field ESS through the communication topology shown in FIG. 9. The PMS (126) of the ESS is bidirectionally connected to the GLS monitoring cloud server (310) via an ETHERNET communication line, and this path is the core communication path of the present invention. The PMS (126) transmits measurement values ​​(p_meas, soc, soh, alarm) to the cloud server (310) at a 1-second interval, and receives command values ​​(p_ref, q_ref, mode) calculated by the cloud server (310) and transmits them to the PCS (122).

[0069] In parallel with this, the power generation company Smart Cloud (320) is operated by being directly connected to the GLS monitoring cloud server (310) via ETHERNET through a separate Smart Gateway (340), and when necessary, it can be connected via ETHERNET to the PMS (126) installed in the GLS ESS product to read power generation, load, and Smart Meter (150) information using Sunspec or HeMS Modbus standards.

[0070] In addition, the user can perform remote monitoring and remote commands by connecting to the GLS monitoring cloud server (310) via WIFI through a Smart Phone (330), and can communicate directly with the PMS (126) in Direct WIFI mode at a location directly adjacent to the ESS. RS485 serial communication is used between the PCS (122) and the PMS (126), and between the PMS (126) and auxiliary devices such as the Smart Meter (150), Meter+EV Charger (152), and AWHP (154).

[0071] Even if communication is interrupted, the previous 24-hour prior schedule (Layer 1 output) is stored inside the PMS (126), so basic operation according to the prior schedule is maintained even without an internet connection. As soon as communication is restored, it is automatically synchronized to the latest schedule through a Rolling Horizon cycle.

[0072] The safety supervisory department (250) passes the corrected command values ​​through the safety gates of SoC (State of Charge), SoH (State of Health), and battery temperature as follows, and then transmits all command values ​​to the PCS via the Modbus TCP protocol.

[0073]

[0074] The safety supervisory department (250) can automatically issue a Safe Idle command (p_ref=0, mode=0) when PCS communication is lost for more than a certain period of time, thereby switching the ESS to a safe stop state.

[0075] The present invention was applied to the applicant's standard model ESS (PCS 60kW · Battery 240kWh, Available 223kWh, Euro efficiency 97%, DoD 97%), and the results of a 24-hour charge-discharge simulation based on the Korea Electric Power Corporation rate system (effective April 16, 2026) are shown in Figure 7. This simulation was based on weekdays in spring and autumn.

[0076] As shown in Figure 7, conventional rule-based (fixed schedule) operation fails to recognize differences in unit prices by time period, so the timing of charging and discharging is not aligned with the optimal unit price range. On the other hand, since the present invention directly incorporates the unit price tariff(t) by time period into the cost function, it automatically converges to a result where charging is concentrated during the off-load hours from 22:00 to 08:00 and discharging is concentrated during the peak load hours from 15:00 to 19:00. As a result, a daily savings of approximately 6,983 won was achieved as shown in Table 1.

[0077]

[0078] The total annual electricity bill savings are calculated by multiplying the daily seasonal savings shown in Table 1 by the number of operating days and summing them. As a result of a simulation conducted by the applicant using 12 months of data at a site where one standard model ESS was applied, the annual comprehensive results shown in Table 2 were obtained.

[0079]

[0080] The effect of the present invention is significantly superior to that of conventional manual rule-based operation. In particular, as seen in the 24-hour simulation (Fig. 7), the algorithm of the present invention (i) automatically concentrates charging during periods of low unit price, (ii) concentrates discharging during periods of high unit price and peak load, and (iii) the daily charge-discharge cycle efficiently utilizes the available battery capacity (223 kWh). This is because the cost function of the present invention explicitly recognizes the asymmetry of unit prices by time period.

[0081] In addition, the same algorithm of the present invention can be applied directly to 60kW commercial ESS and 10kW residential ESS, and when 2 to 10 units are installed in parallel, the site rating is converted to n times and operated with the same algorithm. Therefore, the present invention is commonly applied to all models of the applicant's GLS ESS model series (model names examples: GLSE-HH30K-P3, HH40K-P3, HH50K-P3, HH60K-P3, etc.).

[0082] The system according to the present invention supports remote monitoring, remote updates, and AI operation mode (dynamic mode) operation through, for example, the applicant's Korean GLS server (https: / / k-gls.com), and can automatically perform energy cost minimization operation in AI operation mode when connected to the internet. When the internet is disconnected, it operates according to a 24-hour pre-calculated schedule immediately prior to the disconnection, and when communication is restored, it immediately returns to the latest schedule through a Rolling Horizon cycle.

[0083] The "AI Dynamic Mode" of the present invention is clearly distinguished from the six existing operating modes (hereinafter "existing rules") applied to the conventional ESS shown in Table 3 as follows. While the existing rules all operate in a fixed manner according to the settings when the user selects a mode and manually sets the parameters for each mode before starting operation, the AI ​​Dynamic Mode of the present invention is fundamentally differentiated in that it automatically applies optimal charging and discharging to the ESS equipped with artificial intelligence simply by the user inputting the site's input variables into the server and selecting the "AI Dynamic Mode."

[0084]

[0085] As shown in Table 3, all six existing modes share the common characteristics of (a) requiring users to manually set the mode and mode-specific parameters in advance, (b) failing to adapt to fluctuations in PV generation and load because they operate according to fixed rules once set, and (c) failing to explicitly reflect the asymmetry of unit prices by time of day, season, and holiday in the KEPCO rate table as a cost function. Consequently, the savings effect depends on the user's setting ability, and the savings effect is limited because it is difficult for general users to know the optimal setting values.

[0086] The key differentiators of the present invention can be summarized in the following three points, as shown in Table 4. First, the present invention provides a new operating mode called "AI Dynamic Mode," which is a new mode distinct from the existing six modes. Second, the user only needs to input site-specific input variables once into the operating server (200) of the present invention and select "AI Dynamic Mode," and no manual settings such as time zones, SOC endpoints, and electricity price thresholds required in conventional rule-based modes are required. Third, the algorithm of the present invention defines electricity cost minimization as an explicit cost function and combines MILP optimization for a 24-hour horizon with a Kalman filter and Rolling Horizon correction to implement a new function called "electricity cost minimization" that existing rules cannot provide.

[0087]

[0088] The operating server (200) of the present invention, from the moment the user selects "AI Dynamic Mode," automatically performs (i) a 24-hour PV load forecast for the next day at 23:30 every day, (ii) a 24-hour MILP optimization at 00:00 every day, (iii) a Rolling Horizon correction at the top of every hour, and (iv) automatically sends commands to the PCS (122) via Modbus TCP at a 1-second interval without any additional operation. As a result, the user obtains cost savings without any operational burden.

[0089] In addition, the AI ​​Dynamic Mode of the present invention does not "replace" the existing six modes but "coexists." That is, the operating server (200) operates according to conventional rules when the user selects one of the six existing modes, and the algorithm of the present invention is activated only when the "AI Dynamic Mode" is selected. This maintains compatibility for sites where internet connectivity is unavailable or for cases where the user explicitly desires a specific mode. Furthermore, since the safety supervision unit (250) monitors the SoC, SoH, and temperature limits equally regardless of which mode is selected, system safety is equally guaranteed in all modes.

[0090] According to another aspect of the present invention, a computer-readable recording medium (e.g., hard disk, CD-ROM, flash memory, etc.) is provided, which has a program that can be executed on an electronic device (e.g., server computer, etc.) for the following steps: a procedure for predicting PV generation amount and load demand; a procedure for calculating a charge / discharge schedule for a 24-hour rolling horizon through MILP by reflecting the unit price of the Korea Electric Power Corporation rate table in a cost function; a procedure for smoothing real-time PV generation amount and load values ​​measured at a fixed period using a Kalman filter; a procedure for correcting the charge / discharge schedule in real-time using a rolling horizon when the deviation between the PV generation amount and load real-time values ​​and the charge / discharge schedule exceeds a fixed threshold; and a procedure for transmitting the corrected command value to a power conversion device (PCS).

[0091] As described above, the ESS AI Dynamic operating system according to a preferred embodiment of the present invention has a significant effect compared to the prior art in that it directly reflects the time-of-day, season-of-weekend, and holiday asymmetric unit prices of the Korea Electric Power Corporation's time-of-day rate table into the cost function of the 24-hour horizon MILP, has a combined structure that smooths the measurement noise with a 1-second period using a Kalman filter that uses Joseph form covariance update and then partially updates the pre-schedule with a Rolling Horizon based on the deviation threshold at every hour, and provides a Zero-touch operating mode in which the artificial intelligence automatically applies charging and discharging without separate manual settings, provided that the user inputs the site input variable only once and selects "AI Dynamic Mode."

[0092] Although the present invention has been described above with reference to limited embodiments and drawings, the present invention is not limited thereto, and it is obvious that various modifications and variations within the scope of the technical concept of the present invention are possible by those skilled in the art to which the present invention belongs. Explanation of the symbols

[0093] 110: PV Array 120: ESS 122: PCS 124: Battery 130: Load 140: KEPCO system 200: Operation Server 210: Prediction Unit 220: Optimization Unit 230: Correction Unit 240: Adapter Department 250: Safety Supervision Department

Claims

Claim 1 A method for automatically operating the charging and discharging of an ESS in a seasonal and hourly rate system environment comprises: (a) a step of predicting PV generation and load demand; (b) a step of calculating a charging and discharging schedule for a 24-hour rolling horizon through Mixed Integer Linear Programming (MILP) by reflecting the unit price of the Korea Electric Power Corporation rate table in a cost function; (c) a step of smoothing real-time PV generation and load values ​​measured at a fixed interval using a Kalman filter; and (d) a step of real-time correction of the charging and discharging schedule to the rolling horizon when the deviation between the real-time PV generation and load values ​​and the charging and discharging schedule exceeds a fixed threshold. and (e) a step of transmitting a corrected command value to a power conversion device (PCS); wherein step (a) acquires a 24-hour PV power generation time series from Korea Meteorological Administration short-term forecast data, predicts the PV power generation and load demand through five physical model steps of outdoor solar radiation, panel incidence angle correction, attenuation due to cloud cover, PV cell temperature correction, and inverter efficiency application, and predicts the load demand time series using a machine learning model, and when a user inputs site-specific input variables—including PV capacity, load pattern identifier, Korea Electric Power Corporation tariff code, ESS model, and number of parallel logs—once to the operating server and selects "AI Dynamic Mode," the operating server automatically repeats the prediction of step (a), 24-hour MILP optimization of step (b), rolling horizon correction of step (d), and PCS command transmission of step (e) without additional manual settings by the user, thereby minimizing electricity costs. Claim 2 delete Claim 3 The ESS AI Dynamic operation method according to claim 1, wherein step (b) further comprises the step of calculating the charge / discharge schedule by reflecting unit prices by time period, season, and holiday—including discount unit prices for spring, autumn, Saturday, Sunday, and public holidays in the Korea Electric Power Corporation rate table, intermediate load reduction of Saturday maximum load, and application of low load before Sunday and public holidays—into a cost function, and performing a heuristic fallback in which, if the solver of the MILP fails to produce a solution within a certain period, time periods when the unit price by time period is lower than the daily average are assigned as the charging section, and time periods when it is higher are assigned as the discharging section. Claim 4 ESS AI Dynamic operation method according to claim 1, wherein step (c) is characterized by smoothing real-time PV load values ​​measured at 1-second intervals using a Kalman filter that uses Joseph form covariance updating. Claim 5 ESS AI Dynamic operation method according to claim 1, wherein step (d) is characterized by re-executing the MILP to correct the charge / discharge schedule when the deviation between the actual energy integral value of the previous hour and the charge / discharge schedule at every hour exceeds a predetermined threshold. Claim 6 ESS AI Dynamic operation method according to claim 1, wherein step (e) is characterized by transmitting the corrected command value to the PCS via the Modbus TCP protocol after passing it through safety gates of SoC (State of Charge), SoH (State of Health), and battery temperature. Claim 7 ESS AI Dynamic operation method according to claim 6, characterized in that if the PCS communication of step (e) is lost for more than a certain period of time, the ESS is automatically switched to a safe stop state by issuing a Safe Idle command (p_ref=0, mode=0). Claim 8 delete Claim 9 A system for automatically operating the charging and discharging of an ESS in a seasonal and hourly rate system environment comprises: a forecasting unit that predicts PV generation and load demand; an optimization unit that calculates a charging and discharging schedule for a 24-hour rolling horizon through Mixed Integer Linear Programming (MILP) by reflecting unit prices from the Korea Electric Power Corporation rate table in a cost function; a correction unit that smooths real-time PV generation and load values ​​measured at a fixed interval using a Kalman filter and corrects the charging and discharging schedule in real-time using the rolling horizon when the deviation between the real-time PV generation and load values ​​and the charging and discharging schedule exceeds a predetermined threshold; and an adapter unit capable of communicating with a PCS via Modbus TCP. An ESS AI Dynamic operating system characterized by comprising: a safety supervision unit that blocks abnormal commands by monitoring SoC (State of Charge), SoH (State of Health), and battery temperature limits; wherein the prediction unit acquires a 24-hour PV power generation time series from Korea Meteorological Administration short-term forecast data, predicts the PV power generation and load demand through five physical model steps including outdoor solar radiation, panel incidence angle correction, attenuation due to cloud cover, PV cell temperature correction, and inverter efficiency application, predicts the load demand time series using a machine learning model, and when a user inputs site-specific input variables—including PV capacity, load pattern identifier, KEPCO tariff code, ESS model, and number of parallel units—once to the operating server and selects "AI Dynamic Mode," the operating server automatically and repeatedly performs the prediction of PV power generation and load demand, 24-hour MILP optimization, rolling horizon correction, and PCS command transmission without any additional manual settings by the user thereafter. Claim 10 An ESS AI Dynamic operating system according to claim 9, wherein each of the above-mentioned parts is configured to execute as an independent process so that the safety supervision part alone can issue a Safe Idle command even in the event of a failure in any one part, and the same algorithm is commonly applied to commercial and residential ESSs with variable PCS and battery capacities, and when multiple units are connected in parallel, the site rating is converted and operated using the same algorithm. Claim 11 delete

Images