Park comprehensive energy supply and demand optimization configuration method based on AI multi-modal prediction

Abstract

The invention relates to the technical field of park comprehensive energy optimization, in particular to an AI multi-modal prediction-based park comprehensive energy supply and demand optimization configuration method, which comprises the following steps of: acquiring multi-modal data of a meteorological time sequence, a historical energy load curve, an equipment operation condition record and a real-time energy market price sequence of a park comprehensive energy system; a standardized multi-modal data cube is formed through space-time alignment and missing value completion processing, the standardized multi-modal data cube is input into an AI multi-modal prediction model to process different data modals in parallel, multi-period cooling, heating and power comprehensive load and distributed photovoltaic and wind power generation power prediction values are output, a multi-time-scale operation optimization model is constructed in combination with real-time energy market prices, and a multi-time-scale operation optimization model is constructed. And solving to obtain a real-time scheduling plan before and within the coverage day, and determining power and output instructions of the energy storage, heat storage, gas internal combustion engine and electric refrigerator. According to the method, multi-source data space-time deviation can be eliminated, the prediction matching degree is improved, and precise and collaborative configuration of park comprehensive energy supply and demand is realized.

Description

Technical Field

[0001] This invention relates to the field of integrated energy optimization technology for industrial parks, and in particular to a method for optimizing the allocation of integrated energy supply and demand in industrial parks based on AI multimodal prediction. Background Technology

[0002] The existing integrated energy system in the park relies on single-modal data to predict energy load and distributed new energy generation. The data used only includes a single type of energy load curve or basic meteorological parameters. The energy system operation optimization model is mostly built on a single time scale, and the dispatching plan is only formulated for a single time period. The output control of various energy equipment and the dispatching of energy storage devices are carried out based on single forecast data and fixed price parameters.

[0003] In existing technical solutions, energy-related data from different sources are not spatiotemporally aligned, missing values ​​cannot be effectively filled, and a standardized multimodal data carrier is not formed. The prediction model can only process a single type of data mode and cannot simultaneously complete multi-dimensional predictions of cooling, heating, and electrical loads, as well as wind and solar power generation. The prediction results deviate from the actual energy supply and demand situation in the park. The single-time-scale optimization model cannot match the dynamic operation requirements of the park's energy system at the day-ahead, intraday, and real-time levels. It does not incorporate real-time energy market price sequences to construct scheduling logic. The output commands of electric energy storage, thermal storage devices, gas internal combustion engines, and electric chillers cannot form a multi-time-period coordinated scheduling scheme, resulting in insufficient rationality and adaptability of energy supply and demand allocation.

[0004] This solution requires completing the spatiotemporal alignment, missing value completion, and standardization of multimodal energy data in the park. It also requires using an AI multimodal prediction model to process multiple data modalities in parallel to achieve multidimensional energy parameter prediction. Furthermore, it requires constructing a multi-timescale operation optimization model that integrates prediction results with real-time energy market prices, and outputting detailed equipment scheduling instructions covering multiple time periods. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for optimizing the allocation of integrated energy supply and demand in industrial parks based on AI multimodal prediction.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for optimizing the allocation of comprehensive energy supply and demand in industrial parks based on AI multimodal prediction, comprising:

[0007] Collect multimodal historical and real-time data of the park's integrated energy system. The multimodal historical and real-time data shall include at least meteorological time series, historical energy load curves, equipment operating condition records, and real-time energy market price series.

[0008] The collected multimodal historical and real-time data are spatiotemporally aligned and missing values ​​are filled in to form a standardized multimodal data cube;

[0009] The standardized multimodal data cube is input into a pre-trained AI multimodal prediction model. The AI ​​multimodal prediction model processes different types of data modes in parallel and outputs the predicted values ​​of the combined cooling, heating, and electricity loads of the park for multiple future time periods, as well as the predicted values ​​of the power generation of distributed photovoltaic and wind power.

[0010] The system receives the load forecast and power generation forecast output by the AI ​​multimodal prediction model, and combines them with the real-time energy market price series to construct a multi-timescale operation optimization model for the park's integrated energy system.

[0011] Solving the multi-timescale operation optimization model yields an optimized scheduling plan covering multiple time periods, including day-ahead, intraday, and real-time. The optimized scheduling plan includes the charging and discharging power of the electric energy storage device, the heat storage and release power of the thermal storage device, and the output commands of the gas internal combustion engine and the electric chiller.

[0012] As a further aspect of the present invention, the spatiotemporal alignment and missing value completion of the collected multimodal historical and real-time data includes:

[0013] Identify the timestamps and data sampling frequencies of the meteorological time series, historical energy load curves, equipment operating condition records, and real-time energy market price series;

[0014] Using a preset unified time base and sampling interval, various types of data with different timestamps and sampling frequencies are resampled and interpolated to align all data in the time dimension.

[0015] Missing data points are detected in the spatiotemporally aligned data sequence. A dual fitting method based on the trend of similar data in adjacent time periods and related modal data is used to generate fill values ​​to fill the missing data points, and finally a standardized multimodal data cube that is continuous in time and correlated in modality is formed.

[0016] As a further aspect of the present invention, the standardized multimodal data cube is input into a pre-trained AI multimodal prediction model, including:

[0017] The AI ​​multimodal prediction model includes meteorological feature extraction branches, load feature extraction branches, equipment operating condition feature extraction branches, and market feature extraction branches, which respectively process the corresponding data in the standardized multimodal data cube;

[0018] The meteorological feature extraction branch extracts meteorological feature vectors for future periods from meteorological time series; the load feature extraction branch extracts load time series feature vectors from historical energy load curves; the equipment condition feature extraction branch extracts equipment status feature vectors from equipment operating condition records; and the market feature extraction branch extracts price fluctuation feature vectors from real-time energy market price series.

[0019] The meteorological feature vector, load time series feature vector, equipment status feature vector, and price fluctuation feature vector are fused and input into the spatiotemporal attention prediction network in the AI ​​multimodal prediction model.

[0020] The spatiotemporal attention prediction network outputs predicted values ​​for the park's cooling load, heating load, electricity load, photovoltaic power generation, and wind power generation for multiple future time periods.

[0021] As a further aspect of the present invention, the load forecast and power generation forecast values ​​output by the AI ​​multimodal prediction model are received, and combined with the real-time energy market price series, a multi-time-scale operation optimization model for the park's integrated energy system is constructed, including:

[0022] The goal is to minimize the total operating cost of the park's integrated energy system, which includes the cost of purchasing electricity from the upper-level power grid, the cost of natural gas fuel, and the cost of equipment operation and maintenance.

[0023] The constraints for constructing the multi-timescale operation optimization model include electric power balance constraints, thermal power balance constraints, cold power balance constraints, electric energy storage device operation constraints, thermal storage device operation constraints, and physical operation upper and lower limits and ramp rate constraints for key energy conversion equipment such as gas internal combustion engines and electric chillers.

[0024] The multi-timescale operation optimization model divides the next day into three timescales: day-ahead, intraday, and real-time. At the day-ahead timescale, it makes decisions on equipment start-up and shutdown and energy transfer over long time scales. At the intraday timescale, it performs rolling optimization of power plans over several hours. At the real-time timescale, it handles power deviation balancing at the minute level.

[0025] As a further aspect of the present invention, solving the multi-timescale operation optimization model yields an optimized scheduling plan covering multiple time periods, including:

[0026] The decomposition and coordination algorithm is used to solve the multi-timescale operation optimization model, decomposing the original problem into a daily main problem and multiple intraday sub-problems;

[0027] Solving the main problem of the day, we obtain the planned charging and discharging power curves of the electric energy storage device, the planned heat storage and release power curves of the thermal storage device, and the planned output curves of the gas internal combustion engine and the electric chiller for the next 24 hours at one-hour intervals.

[0028] Based on the solution results of the main problem of the day, the intraday sub-problems covering the next few hours are solved in a rolling manner, and the planned charging and discharging power curves of the energy storage equipment, the planned heat storage and release power curves of the thermal storage device, and the planned output curves of the key equipment are refined and corrected to form an intraday rolling plan with a 15-minute interval.

[0029] The solution to the day-ahead master problem and the intraday rolling plan together constitute an optimized scheduling plan.

[0030] As a further aspect of the present invention, it also includes:

[0031] The daytime schedule in the optimized scheduling plan is decomposed into specific equipment start-up and shutdown instructions and baseline power curves, and then sent to each energy conversion and storage device for execution.

[0032] During the intraday execution phase, based on the latest ultra-short-term load forecast and power generation forecast, the intraday rolling plan in the optimized scheduling plan is revised to generate equipment power adjustment instructions within the current time window;

[0033] During the real-time execution phase, the actual operating power of each energy conversion and storage device is monitored and compared with the device power adjustment command. Power compensation signals are generated through real-time feedback control.

[0034] The device start / stop commands, reference power curves, device power adjustment commands, and power compensation signals are aggregated to form the final control command set for each energy device;

[0035] Based on the final set of control instructions for each energy device, the physical devices in the park's integrated energy system are driven to perform corresponding energy production, conversion, storage, and consumption actions, thereby completing the optimal allocation of supply and demand.

[0036] The step of decomposing the day-ahead plan in the optimized scheduling plan into specific equipment start / stop instructions and baseline power curves includes:

[0037] The planned charge and discharge power curves of the energy storage devices in the day-ahead portion of the optimized scheduling plan are analyzed to generate charge and discharge status commands and baseline charge and discharge power values ​​for the energy storage devices at the start of each hour.

[0038] The thermal storage and release power curves of the thermal storage devices in the day-ahead portion of the optimized scheduling plan are analyzed to generate thermal storage or release status commands and reference power values ​​for the thermal storage devices at the start of each hour.

[0039] The planned output curves of the gas internal combustion engine and electric chiller in the daytime portion of the optimized scheduling plan are analyzed to generate start-stop status commands and baseline operating power values ​​for the gas internal combustion engine and electric chiller within each hour.

[0040] As a further aspect of the present invention, during the intraday execution phase, the intraday rolling plan in the optimized scheduling plan is revised based on the latest ultra-short-term load forecast and power generation forecast, including:

[0041] During the intraday execution phase, the AI ​​multimodal prediction model is invoked to perform ultra-short-term predictions, obtaining updated cold, heat, and electricity load forecasts and photovoltaic and wind power generation forecasts for the next few hours.

[0042] The relevant parameters in the multi-timescale operation optimization model are updated with the latest ultra-short-term forecast values, and the intraday rolling plan for the next few hours is re-solved to obtain the updated equipment power plan curve.

[0043] Compare the updated equipment power plan curve with the intraday rolling plan in the original optimized scheduling plan, and calculate the power adjustment amount for each equipment;

[0044] Based on the power adjustment amount of each device, generate equipment power adjustment instructions for electric energy storage devices, thermal storage devices, gas internal combustion engines, and electric chillers within the current time window.

[0045] As a further aspect of the present invention, the step of monitoring the actual operating power of each energy conversion and storage device during the real-time execution phase, comparing it with the device power adjustment command, and generating a power compensation signal through real-time feedback control includes:

[0046] Real-time acquisition of the actual output power or storage power of electric energy storage devices, thermal storage devices, gas internal combustion engines and electric chillers;

[0047] The actual power of each device is compared with the power setting value required in the device power adjustment command at the current time, and the power deviation value is calculated.

[0048] The power deviation value is input to a proportional-integral-derivative (PID) controller, which outputs a power compensation signal to eliminate the power deviation.

[0049] As a further aspect of the present invention, the device start / stop commands, reference power curves, device power adjustment commands, and power compensation signals are aggregated to form a final set of control commands for each energy device, including:

[0050] For an energy storage device, the charging and discharging status command in the device start / stop command, the reference charging and discharging power value in the reference power curve, the power adjustment amount in the device power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command of the energy storage device.

[0051] For a thermal storage device, the thermal storage and release status command in the device start / stop command, the reference power value in the reference power curve, the power adjustment amount in the device power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command of the thermal storage device.

[0052] For both gas internal combustion engines and electric chillers, the start / stop status command in the equipment start / stop command, the reference operating power value in the reference power curve, the power adjustment amount in the equipment power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command for the gas internal combustion engine and the electric chiller.

[0053] As a further aspect of the present invention, based on the final set of control instructions for each energy device, the physical devices in the integrated energy system of the park are driven to perform corresponding energy production, conversion, storage, and consumption actions to complete the optimized supply and demand configuration, including:

[0054] The final instantaneous power control command of the energy storage device is converted into a drive signal of the power electronic converter to control the charging or discharging process of the energy storage battery.

[0055] The final instantaneous power control command of the heat storage device is converted into water pump and valve opening adjustment signals to control the heat storage or heat release process of the heat storage tank.

[0056] The final instantaneous power control command of the gas internal combustion engine is converted into a fuel supply regulation signal and a generator excitation regulation signal to control the output electric power of the gas internal combustion engine.

[0057] The final instantaneous power control command of the electric chiller is converted into a compressor frequency adjustment signal to control the cooling power output of the electric chiller.

[0058] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0059] The system performs spatiotemporal alignment and missing value completion on meteorological time series, historical energy load curves, equipment operating condition records, and real-time energy market price series to form a standardized multimodal data cube. This data cube is then input into a pre-trained AI multimodal prediction model. The model can process different types of data modalities in parallel and simultaneously output the predicted values ​​of the combined cooling, heating, and electricity loads of the park for multiple future time periods, as well as the predicted values ​​of distributed photovoltaic and wind power generation. The spatiotemporal alignment of the multimodal data eliminates the spatiotemporal dimensional differences between different data sources, and missing value completion maintains the integrity of the data sequence. The standardized data cube provides a regular input carrier for the model, and the multimodal parallel processing adapts to the characteristic logic of different types of data. The multi-time period and multi-type prediction outputs fit the actual operating characteristics of the park's integrated energy system.

[0060] Based on the load and power generation forecasts output by the AI ​​multimodal prediction model, a multi-timescale operation optimization model for the park's integrated energy system is constructed by combining real-time energy market price series. Solving this optimization model yields optimized scheduling plans covering multiple time periods, including day-ahead, intraday, and real-time. The model clarifies the charging and discharging power of energy storage devices, the heat storage and release power of thermal storage devices, and the output commands of gas internal combustion engines and electric chillers. The multi-timescale optimization model matches the operational variation characteristics of the park's energy system at different times. The integration of real-time energy market price series allows the scheduling plan to adapt to market-based trading rules. The specific power and output commands of various types of equipment refine the scheduling execution parameters. The multi-time-period scheduling plan forms a continuous control scheme, ensuring that the supply and demand configuration of the park's integrated energy system aligns with real-time operation and market trading needs. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method for optimizing the allocation of integrated energy supply and demand in a park based on AI multimodal prediction, as described in this invention.

[0062] Figure 2 A flowchart for constructing a multi-timescale operation optimization model for the park's integrated energy system;

[0063] Figure 3 This is a breakdown diagram of the park's day-ahead energy plan;

[0064] Figure 4 Real-time PID feedback control diagram for the park's energy system;

[0065] Figure 5 This is a schematic diagram of the synthesis of power control commands across multiple time scales and the thermal storage status of the thermal storage device in the park. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0067] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0068] See Figure 1 This invention provides a method for optimizing the supply and demand of integrated energy in a park based on AI multimodal prediction. The method includes: collecting multimodal historical and real-time data of the park's integrated energy system, including meteorological time series, historical energy load curves, equipment operating condition records, and real-time energy market price series. The collected multimodal historical and real-time data are spatiotemporally aligned and missing values ​​are filled in to form a standardized multimodal data cube. The standardized multimodal data cube is input into a pre-trained AI multimodal prediction model, which processes different types of data modes in parallel and outputs predicted values ​​for the integrated cooling, heating, and electricity loads of the park for multiple future time periods, as well as predicted values ​​for the power generation of distributed photovoltaic and wind power. The load and power generation prediction values ​​output by the AI ​​multimodal prediction model are received, and combined with the real-time energy market price series, a multi-timescale operation optimization model for the park's integrated energy system is constructed. The multi-timescale operation optimization model is solved to obtain an optimized scheduling plan covering multiple time periods, including day-ahead, intraday, and real-time. This plan includes the charging and discharging power of electric energy storage devices, the heat storage and release power of thermal storage devices, and the output commands of gas internal combustion engines and electric chillers.

[0069] In one embodiment of the present invention, the timestamps and data sampling frequencies of meteorological time series, historical energy load curves, equipment operating condition records, and real-time energy market price series are identified. Using a preset unified time base and sampling interval, various types of data with different timestamps and sampling frequencies are resampled and interpolated to align all data in the time dimension. Missing data points in the spatiotemporally aligned data series are detected, and a dual fitting method based on the trends of similar data in adjacent time periods and related modal data is used to generate filler values ​​to fill the missing data points, ultimately forming a standardized multimodal data cube that is continuous in time and correlated in modality. The AI ​​multimodal prediction model includes meteorological feature extraction branches, load feature extraction branches, equipment operating condition feature extraction branches, and market feature extraction branches, which respectively process the corresponding data in the standardized multimodal data cube. The meteorological feature extraction branch extracts meteorological feature vectors for future periods from the meteorological time series; the load feature extraction branch extracts load time series feature vectors from the historical energy load curves; the equipment operating condition feature extraction branch extracts equipment status feature vectors from equipment operating condition records; and the market feature extraction branch extracts price fluctuation feature vectors from the real-time energy market price series. The meteorological feature vector, load time series feature vector, equipment status feature vector, and price fluctuation feature vector are fused and input into the spatiotemporal attention prediction network in the AI ​​multimodal prediction model. The spatiotemporal attention prediction network outputs predicted values ​​for the park's cooling load, heating load, electricity load, photovoltaic power generation, and wind power generation for multiple future time periods.

[0070] In specific implementations, meteorological time series data may include hourly observations of temperature, humidity, irradiance, and wind speed; historical energy load curves may include 15-minute power data for cooling, heating, and electrical loads; equipment operating condition records may include minute-level logs of start-up and shutdown times, operating power, and efficiency of gas-fired internal combustion engines; and real-time energy market price series may include 15-minute data on time-of-use electricity prices. It is understood that these data sources have different timestamps and sampling frequencies. The timestamp for the meteorological time series may be the hour, with a sampling frequency of once per hour; the timestamp for the historical energy load curve is the start time every 15 minutes, with a sampling frequency of once every 15 minutes; the timestamp for the equipment operating condition record is the trigger time for changes in equipment status, with a variable sampling frequency; and the timestamp for the real-time energy market price series is synchronized with the electricity market settlement cycle, with a sampling frequency of once every 15 minutes. In some embodiments, a spatiotemporal alignment operation is performed, setting a unified time base to Coordinated Universal Time (UTC) time zone zero and a unified sampling interval of 15 minutes. For the meteorological time series, a linear interpolation method is used to generate a meteorological data sequence with 15-minute intervals between each retained hourly meteorological data point. The original temperature at 12:00 is 25 degrees Celsius, and the temperature at 13:00 is 26 degrees Celsius. The linear interpolation yields temperatures of 25.25 degrees Celsius at 12:15, 25.5 degrees Celsius at 12:30, and 25.75 degrees Celsius at 12:45. For the historical energy load curve, the original sampling interval is already 15 minutes, so no resampling interpolation is required. For equipment operating condition records, the average operating power of the equipment within a 15-minute window is calculated as the equipment status value at the end of that period. For example, within the window from 12:00 to 12:15, if a gas-fired internal combustion engine starts at 12:05 and runs at 500 kW until 12:10 before shutting down, the average operating power of the equipment in that window is approximately (500 kW * 5 minutes) / 15 minutes ≈ 166.67 kW. The real-time energy market price series itself is 15-minute data, so its values ​​are used directly. After the above resampling and interpolation, the data of all modalities are aligned in the time dimension, forming a data series with a unified timestamp.

[0071] In practical implementation, missing points in the aligned data sequence are detected. At a certain point in time, the irradiance data of the meteorological time series is missing due to sensor failure, while the temperature and humidity data at that moment are complete, and the electrical load data in the historical energy load curve is also complete. A dual fitting method based on the trends of similar data in adjacent time periods and related modal data is used for completion. On the one hand, the average value of similar data in the time period before and after the missing moment, i.e., the irradiance data, is calculated as the basic completion value. On the other hand, the trend of related modal data is analyzed. Since irradiance is strongly correlated with photovoltaic power generation, and although the historical energy load curve does not contain photovoltaic power, the correlation pattern between irradiance and temperature and humidity on historical days at the same time can be analyzed to establish a simple mapping relationship. Combining the two, the final completion value is generated. Optionally, the dual fitting process can be implemented through a weighted formula as follows:

[0072]

[0073] in: This represents the irradiance complement at time t. and Let represent the irradiance observations for the two adjacent time periods before and after time t, respectively. and These represent the temperature and humidity values ​​at time t, respectively. This represents the historical daily average background irradiance value for the same hour as time t. This represents a description learned from historical data. The mapping function to the irradiance estimate, It is a fusion coefficient between 0 and 1, used to balance the weights of adjacent data and related trends. It can be understood that by processing all missing points, a standardized multimodal data cube that is continuous in time and correlated in modality is finally formed, whose dimensions can be represented as [time step, number of modalities, feature dimension].

[0074] In its implementation, the AI ​​multimodal prediction model comprises four parallel feature extraction branches, each processing different modalities of data. The meteorological feature extraction branch takes aligned meteorological time series as input and may consist of a convolutional neural network and a long short-term memory network, extracting meteorological feature vectors representing future meteorological evolution patterns from the input sequence. The load feature extraction branch takes historical energy load curves as input and may consist of an attention-enhanced long short-term memory network, extracting load time-series feature vectors from historical load changes. The equipment operating condition feature extraction branch takes aligned sequences of equipment operating condition records as input and may consist of fully connected layers, extracting equipment state feature vectors reflecting equipment health and operational constraints. The market feature extraction branch takes real-time energy market price sequences as input and may consist of a one-dimensional convolutional network, extracting price fluctuation feature vectors.

[0075] In some embodiments, meteorological feature vectors, load time-series feature vectors, equipment status feature vectors, and price fluctuation feature vectors from four branches are concatenated or weighted and fused into a unified multimodal fusion feature vector. Optionally, a modal attention mechanism can be introduced into the fusion process to dynamically assign importance weights to different modal features. This unified multimodal fusion feature vector is then input into the spatiotemporal attention prediction network in the AI ​​multimodal prediction model. The spatiotemporal attention prediction network typically consists of an encoder-decoder architecture and embeds an attention mechanism to capture the spatiotemporal dependencies between sequences. The spatiotemporal attention prediction network ultimately outputs multiple future time periods, such as the predicted values ​​for park cooling load, heating load, electricity load, photovoltaic power generation, and wind power generation at 15-minute intervals over the next 24 hours. Each predicted value is a power value sequence corresponding to a future time.

[0076] In one embodiment of the present invention, see [reference] Figure 2The goal is to minimize the total operating cost of the park's integrated energy system, which includes the cost of purchasing electricity from the upper-level power grid, the cost of gas fuel, and the cost of equipment operation and maintenance. Constraints are established for a multi-timescale operation optimization model, including constraints on power balance, thermal power balance, cooling power balance, operation of electric energy storage equipment, operation of thermal storage devices, and physical operating limits and ramp-up rates for key energy conversion equipment such as gas internal combustion engines and electric chillers. The multi-timescale operation optimization model divides the next day into three timescales: day-ahead, intraday, and real-time. At the day-ahead scale, decisions are made regarding equipment start-up and shutdown and long-term energy transfer. At the intraday scale, rolling optimization of power plans is performed over several hours. At the real-time scale, minute-level power deviation balancing is handled. A decomposition and coordination algorithm is used to solve the multi-timescale operation optimization model, decomposing the original problem into a day-ahead master problem and multiple intraday sub-problems. Solving the day-ahead master problem yields the planned charge / discharge power curves for electric energy storage equipment, the planned heat storage / release power curves for thermal storage devices, and the planned output curves for gas internal combustion engines and electric chillers for the next 24 hours at one-hour intervals. Based on the solution to the main problem, a rolling solution is performed on sub-problems covering the next few hours. This refines and corrects the planned charging and discharging power curves of energy storage devices, the planned heat storage and release power curves of thermal storage devices, and the planned output curves of key equipment, forming a rolling plan with 15-minute intervals. The solution to the main problem and the rolling plan together constitute an optimized scheduling plan.

[0077] In the specific implementation, the process of constructing and solving the multi-timescale operation optimization model is described. The following detailed explanation is provided with examples and data comparisons. In the specific implementation, the system receives the predicted sequences of cooling, heating, and electrical loads for the next 24 hours at a time resolution of 15 minutes, as well as the predicted sequences of photovoltaic and wind power generation, output by the AI ​​multimodal prediction model. Simultaneously, it combines these with real-time energy market price sequences at the same time scale to construct a multi-timescale operation optimization model for the park's integrated energy system. The model aims to minimize the total operating cost of the park's integrated energy system, which includes the cost of purchasing electricity from the upper-level grid, the cost of natural gas fuel, and the cost of equipment operation and maintenance. In the specific implementation, the objective function can be expressed by the following formula:

[0078]

[0079] in: This represents the total operating cost during the optimization period. Index representing time interval, This represents the total number of time intervals within the optimization period. This represents the real-time electricity price for electricity purchased from the upstream power grid within a time interval t. This represents the average electrical power purchased from the upstream power grid during time interval t. This represents the duration of each time interval, therefore The term represents the cost of purchasing electricity during the time interval t. Indicates the price of gas. This represents the volume of gas consumed by the gas-fired internal combustion engine during time interval t. This represents the unit power operation and maintenance cost coefficient of the i-th energy conversion or storage device. This represents the average output power or storage power of the i-th device during the time interval t. This represents the maintenance cost of the equipment over time interval t. This is understandable. From real-time energy market price series, There is a conversion relationship between the power generation capacity and power generation efficiency of the gas internal combustion engine and the equipment operation and maintenance cost, which is the operating wear and tear cost of the main equipment such as gas internal combustion engines, electric chillers, electric energy storage devices, and thermal storage devices.

[0080] In practical implementation, constraints are constructed for the multi-timescale operation optimization model, including power balance constraints, thermal power balance constraints, and cooling power balance constraints. In the power balance constraint, for any time interval t, the average power purchased from the upstream power grid... Average power generation of gas-fired internal combustion engines Average power of photovoltaic power generation Average power of wind power generation and the average discharge power of energy storage devices The sum must equal the average power of the electrical load. Average electrical power consumed by electric chillers and the average charging power of energy storage devices The sum of Under the thermal power balance constraint, for any time interval t, the average waste heat power generated by the gas internal combustion engine is... Average heat release power of the heat storage device The sum must equal the average power of the heat load. Average power of thermal storage device The sum of In the cooling power balance constraint, for any time interval t, the average cooling power generated by the electric chiller is... It must be equal to the average power of the cooling load. ,Right now In specific implementation, the operational constraints of electric energy storage devices include: the state of charge (SOC) of the electric energy storage device must be within its rated capacity at any time interval t; the charging and discharging power of the electric energy storage device cannot exceed its maximum allowable power, and charging and discharging cannot occur simultaneously; the SOC of the electric energy storage device must be set consistently at the beginning and end of the optimization cycle. The operational constraints of thermal storage devices include: the heat storage capacity of the thermal storage device must be within its rated capacity at any time interval t; the heat storage and heat release power of the thermal storage device cannot exceed its maximum allowable power, and heat storage and heat release cannot occur simultaneously. The physical operational constraints of key energy conversion equipment such as gas internal combustion engines and electric chillers include: the output electrical power of the gas internal combustion engine must be between its minimum technical output and rated power; the power change of the gas internal combustion engine between two adjacent time intervals cannot exceed its maximum ramp rate; the cooling power of the electric chiller must be between its minimum technical output and rated power; the power change of the electric chiller between two adjacent time intervals cannot exceed its maximum ramp rate. Optionally, the operational constraints of the gas internal combustion engine also need to consider its minimum start-stop time limit.

[0081] In some embodiments, the multi-timescale operation optimization model divides the future day into three timescales: day-ahead, intraday, and real-time. At the day-ahead scale, decisions are made regarding equipment start-up and shutdown, and long-term energy transfer. The optimization time interval is typically set to 1 hour. Decision variables include the start-up and shutdown status of the gas internal combustion engine, the average hourly charging and discharging power plan of the electric energy storage device, and the average hourly heat storage and dissipation power plan of the thermal storage device. At the intraday scale, rolling optimization of the power plan is performed over several hours, with an optimization time interval set to 15 minutes, refining the equipment power within the day-ahead planning framework. At the real-time scale, minute-level power deviation balancing is handled, with time intervals of 5 minutes or less, using feedback control to compensate for prediction errors and actual equipment operating deviations. In specific implementations, the day-ahead optimization model is solved based on 24-hour forecast data, the intraday rolling optimization model is solved based on ultra-short-term forecast data for the next 4 to 6 hours, and real-time balancing is performed online based on actual measurement data.

[0082] In practical implementation, a decomposition and coordination algorithm is used to solve the multi-timescale operation optimization model, decomposing the original problem into a day-ahead master problem and multiple intraday sub-problems. Solving the day-ahead master problem yields the planned charging and discharging power curves of the electric energy storage device, the planned heat storage and release power curves of the thermal storage device, and the planned output curves of the gas internal combustion engine and electric chiller for the next 24 hours at one-hour intervals. For example, the solution to the day-ahead master problem might show that during the peak electricity price period of 13:00-14:00, the planned discharge power of the electric energy storage device is 300 kW, and the planned output of the gas internal combustion engine is 800 kW; during the off-peak electricity price period of 02:00-03:00, the planned charging power of the electric energy storage device is 250 kW, and the gas internal combustion engine is planned to be shut down. Based on the solution to the day-ahead master problem, intraday sub-problems covering the next few hours are solved in a rolling manner. In some embodiments, at 10:00 AM each day, based on the daily schedule from 10:00 AM to 10:00 AM the following day, a rolling solution is performed for intraday sub-problems covering the next four hours, i.e., from 10:00 AM to 2:00 PM. These intraday sub-problems are refined and corrected at 15-minute intervals to detail the planned charge / discharge power curves of the energy storage devices, the planned heat storage / discharge power curves of the thermal storage devices, and the planned output curves of key equipment. For example, for the hour 12:00 PM to 1:00 PM, if the daily schedule gives an average discharge power of 300 kW for the energy storage devices, the intraday sub-problem might refine it to: 280 kW discharge from 12:00 PM to 12:15 PM, 320 kW discharge from 12:15 PM to 12:30 PM, 310 kW discharge from 12:30 PM to 12:45 PM, and 290 kW discharge from 12:45 PM to 1:00 PM. The average for the four periods remains 300 kW, but the solution provides a more refined response to ultra-short-term load fluctuations. It is understandable that the solution to the day-ahead master problem, together with the intraday rolling plan, constitutes the optimized scheduling plan from the day-ahead to the intraday scale. Optionally, the decomposition coordination algorithm passes the hourly end-of-hour state of charge of the energy storage devices and the hourly end-of-hour heat storage of the thermal storage devices as coupling variables to the intraday sub-problems in the day-ahead master problem, and feeds back the optimized device power sequence of the intraday sub-problems to the day-ahead master problem for iterative coordination until the plan converges.

[0083] In one embodiment of the present invention, the day-ahead schedule is decomposed into specific equipment start-up and shutdown commands and reference power curves, which are then issued to each energy conversion and storage device for execution. The planned charge-discharge power curves of the electric energy storage devices in the day-ahead portion of the optimized scheduling plan are analyzed to generate charge-discharge status commands and reference charge-discharge power values ​​for the electric energy storage devices at the start of each hour. The planned heat storage and release power curves of the thermal storage devices in the day-ahead portion of the optimized scheduling plan are analyzed to generate heat storage or release status commands and reference power values ​​for the thermal storage devices at the start of each hour. The planned output curves of the gas internal combustion engine and electric chiller in the day-ahead portion of the optimized scheduling plan are analyzed to generate start-up and shutdown status commands and reference operating power values ​​for the gas internal combustion engine and electric chiller within each hour. During the intraday execution phase, based on the latest ultra-short-term load forecast and power generation forecast, the intraday rolling plan in the optimized scheduling plan is revised to generate equipment power adjustment commands within the current time window. During the real-time execution phase, the actual operating power of each energy conversion and storage device is monitored and compared with the equipment power adjustment commands, and power compensation signals are generated through real-time feedback control. The system aggregates equipment start / stop commands, reference power curves, equipment power adjustment commands, and power compensation signals to form the final control command set for each energy device. Based on this final control command set, the system drives the physical equipment in the park's integrated energy system to perform corresponding energy production, conversion, storage, and consumption actions, thereby achieving optimized supply and demand allocation.

[0084] In practical implementation, after obtaining the optimized scheduling plan generated by the above embodiment, the day-ahead plan in the optimized scheduling plan is decomposed into specific equipment start / stop commands and baseline power curves, and then sent to each energy conversion and storage device for execution. The planned charge / discharge power curves of the electric energy storage devices in the day-ahead portion of the optimized scheduling plan are analyzed to generate charge / discharge status commands and baseline charge / discharge power values ​​for the electric energy storage devices at the start of each hour. The planned heat storage / release power curves of the thermal storage devices in the day-ahead portion of the optimized scheduling plan are analyzed to generate heat storage or heat release status commands and baseline power values ​​for the thermal storage devices at the start of each hour. The planned output curves of the gas internal combustion engine and electric chiller in the day-ahead portion of the optimized scheduling plan are analyzed to generate start / stop status commands and baseline operating power values ​​for the gas internal combustion engine and electric chiller within each hour. As can be understood, the decomposition process transforms a continuous power curve into discrete time-point commands and constant values. For example, for an energy storage device with a rated capacity of 500 kWh, the planned charge-discharge power curve for the period from 10:00 to 11:00 in the optimized scheduling plan shows an average discharge power of 50 kW. After decomposition, the command generated at 10:00:00 is "discharge state," with a baseline charge-discharge power value of -50 kW (the negative sign indicates discharge). Refer to Table 1, which shows a simplified command decomposition fragment.

[0085] Table 1: Optimized Scheduling Plan Decomposition Table

[0086] In some embodiments, during the intraday execution phase, the intraday rolling plan in the optimized scheduling plan is revised based on the latest ultra-short-term load and power generation forecasts, generating equipment power adjustment instructions for the current time window. At 10:15 daily, the AI ​​multimodal prediction model is invoked to execute ultra-short-term forecasts, obtaining updated cold, heat, and electricity load forecasts and photovoltaic and wind power generation forecasts for the next 4 hours (10:15 to 14:15). The relevant parameters in the multi-timescale running optimization model are updated with the latest ultra-short-term forecast values, and the intraday rolling plan for the next few hours is re-solved, resulting in an updated equipment power plan curve. The updated equipment power plan curve is compared with the original intraday rolling plan in the optimized scheduling plan, and the power adjustment amount for each device is calculated. For example, for the 15-minute period from 10:15 to 10:30, the planned power of the energy storage device in the original intraday rolling plan was 55 kW for discharging, while in the updated plan based on the latest forecast, the planned power of the device becomes 10 kW for charging. The power adjustment amount is... for Based on the power adjustment amount of each device, power adjustment instructions are generated for the electric energy storage device, thermal storage device, gas internal combustion engine, and electric chiller within the current time window. Optionally, the power adjustment amount ΔP can be calculated using a unified expression:

[0087]

[0088] in: This indicates the amount of power adjustment for a specific device at a specific future time period. This represents the planned power value of the device for that time period in the updated schedule based on the latest ultra-short-term forecast solution. This indicates the planned power output of the equipment during the same time period in the original optimized scheduling plan. A positive value indicates that the output or charging power needs to be increased, while a negative value indicates that the output needs to be reduced or the equipment needs to be switched to discharging / heat release.

[0089] In practical implementation, during the real-time execution phase, the actual operating power of each energy conversion and storage device is monitored and compared with the device power adjustment command. A power compensation signal is generated through real-time feedback control. The actual output power or storage power of the electric energy storage device, thermal storage device, gas internal combustion engine, and electric chiller are collected in real time. The actual power of each device is compared with the power setpoint required in the device power adjustment command at the current moment to calculate the power deviation value. For example, at 10:20, the device power adjustment command for the electric energy storage device requires an output charging power of 10 kW, while the actual collected power is 8 kW, resulting in a power deviation value of -2 kW. The power deviation value is input to the proportional-integral-derivative (PID) controller, which outputs a power compensation signal to eliminate the power deviation. The output signal u(t) of the PID controller is calculated based on the power deviation e(t).

[0090]

[0091] in: , , These are the proportional, integral, and differential coefficients, respectively. This is the real-time power deviation value. This is the generated power compensation signal.

[0092] In practice, the equipment start / stop commands, reference power curves, equipment power adjustment commands, and power compensation signals are aggregated to form the final control command set for each energy device. This can be understood as a process of command superposition. For energy storage devices, the charging / discharging status command from the equipment start / stop command, the reference charging / discharging power value from the reference power curve, the power adjustment amount from the equipment power adjustment command, and the power compensation value from the power compensation signal are superimposed to obtain the final instantaneous power control command for the energy storage device. For example, at 10:20, the equipment start / stop command requires a "charging" state, the reference power curve gives a reference charging / discharging power value of -50 kW, the intraday correction-generated equipment power adjustment command requires an increase of 65 kW, and the real-time feedback-generated power compensation signal is +2 kW. Therefore, the final instantaneous power control command is: Charging, power value = (-50) + 65 + 2 = 17 kW. For thermal storage devices, the final instantaneous power control command is obtained by superimposing the thermal storage / release status command from the device start / stop command, the reference power value from the reference power curve, the power adjustment amount from the device power adjustment command, and the power compensation value from the power compensation signal. For gas internal combustion engines and electric chillers, the final instantaneous power control command is obtained by superimposing the start / stop status command from the device start / stop command, the reference operating power value from the reference power curve, the power adjustment amount from the device power adjustment command, and the power compensation value from the power compensation signal. In some embodiments, the status command has the highest priority; if the superimposed power command symbol conflicts with the status command, the power symbol is adjusted according to the status command.

[0093] In practical implementation, based on the final control command set for each energy device, the physical devices in the park's integrated energy system are driven to perform corresponding energy production, conversion, storage, and consumption actions, achieving optimized supply and demand configuration. The final instantaneous power control command for the electric energy storage device is converted into a drive signal for the power electronic converter, controlling the charging or discharging process of the electric energy storage battery. For example, a 17 kW charging command is converted into the converter's pulse width modulation signal duty cycle. The final instantaneous power control command for the thermal storage device is converted into pump and valve opening adjustment signals, controlling the thermal storage or release process of the thermal storage tank. For example, a positive power command might correspond to increasing the hot water pump frequency and adjusting the valve opening to increase the thermal storage flow. The final instantaneous power control command for the gas internal combustion engine is converted into a fuel supply adjustment signal and a generator excitation adjustment signal, controlling the output power of the gas internal combustion engine. The final instantaneous power control command for the electric chiller is converted into a compressor frequency adjustment signal, controlling the cooling power output of the electric chiller. Optionally, these conversion processes are implemented through the local programmable logic controller or dedicated controller of each device.

[0094] See Figure 3This is a breakdown of the park's energy day-ahead plan, fully presenting the baseline power curves of the four core energy devices. Electric energy storage devices fluctuate between -60kW and +80kW, with negative values ​​indicating discharge and power output to the park's grid, while positive values ​​indicate charging and energy absorption from the grid. Thermal storage devices fluctuate between -30kW and +50kW, with positive values ​​indicating heat storage and heat release to meet the park's heat load. Gas turbines, operating at 400kW to 500kW, are continuously positive and operate without shutdown, supplying the park's basic electricity and heat loads. Their power dynamically adjusts according to the park's energy demand to ensure stable system operation. Electric chillers fluctuate with the park's cooling load demand, ensuring cooling needs in offices and commercial areas, and work in conjunction with the gas turbines to achieve combined cooling, heating, and power (CCHP). The diagram clearly shows the power synergy between electric energy storage, thermal storage, gas turbines, and electric chillers, reflecting the multi-energy complementarity and integrated source-grid-load-storage characteristics of the park's comprehensive energy system.

[0095] In one embodiment of the present invention, during the intraday execution phase, an AI multimodal prediction model is invoked to perform ultra-short-term prediction, obtaining updated cold, heat, and electricity load forecasts and photovoltaic and wind power generation forecasts for the next few hours. The relevant parameters in the multi-timescale operation optimization model are updated with the latest ultra-short-term forecast values, and the intraday rolling plan for the next few hours is re-solved to obtain an updated equipment power plan curve. The updated equipment power plan curve is compared with the intraday rolling plan in the original optimized scheduling plan to calculate the power adjustment amount for each device. Based on the power adjustment amount for each device, equipment power adjustment instructions for the electric energy storage device, thermal storage device, gas internal combustion engine, and electric chiller within the current time window are generated. The actual output power or stored power of the electric energy storage device, thermal storage device, gas internal combustion engine, and electric chiller are collected in real time. The actual power of each device is compared with the power setpoint required in the equipment power adjustment instruction corresponding to the current time to calculate the power deviation value. The power deviation value is input to the proportional-integral-derivative (PID) controller, which outputs a power compensation signal to eliminate the power deviation.

[0096] In practice, during the scheduled timeframes of the daily execution phase, such as the start of each rolling optimization window at 10:15, 10:30, and 10:45 each day, the AI ​​multimodal prediction model is invoked to perform ultra-short-term forecasts. The AI ​​multimodal prediction model receives the latest multimodal data cube up to the current moment and executes the forecast process to obtain updated cold, heat, and electricity load forecasts, as well as photovoltaic and wind power generation forecasts for the next few hours. For example, at 10:15, ultra-short-term forecasting yields updated forecasts for the next four hours, from 10:15 to 14:15, at 15-minute intervals for a total of 16 time points. It can be understood that the updated electricity load forecast might show a slightly higher load level in the next hour than previously predicted, while the updated photovoltaic power generation forecast might be lower than previously predicted due to cloud cover changes.

[0097] In practical implementation, the relevant parameters in the multi-timescale operation optimization model are updated with the latest ultra-short-term forecast values. Specifically, the cooling load, heating load, and electrical load parameters corresponding to the future rolling window in the optimization model are replaced with the updated cooling load forecast values, heating load forecast values, and electrical load forecast values. The photovoltaic power generation and wind power generation parameters are also replaced with the updated photovoltaic power generation forecast values ​​and wind power generation forecast values. Within the framework of the intraday sub-problem, the intraday rolling plan for the next few hours is resolved to obtain the updated equipment power plan curve. The updated equipment power plan curve is compared with the intraday rolling plan in the original optimized scheduling plan to calculate the power adjustment amount for each device in each future time period. Refer to Table 2, which shows a comparison of the power plans and adjustment amount calculations for some devices within the 10:30 to 11:00 time window.

[0098] Table 2: Intraday Rolling Plan Revision Table

[0099] In some embodiments, based on the power adjustment amount of each device, power adjustment instructions for the electric energy storage device, thermal storage device, gas internal combustion engine, and electric chiller within the current time window are generated. Each device power adjustment instruction is a sequence of power adjustment values ​​for each controlled device, covering several future time periods. For example, for the data in Table 2, the generated device power adjustment instructions would include specific instructions such as "at 10:30:00, increase the power setpoint of the electric energy storage device by 50 kW" and "at 10:30:00, increase the power setpoint of the gas internal combustion engine by 20 kW". Optionally, the device power adjustment instructions are sent to the local controller in the form of an instruction list or a time-series data packet.

[0100] In practical implementation, during the real-time execution phase, the actual operating power of each energy conversion and storage device is monitored and compared with the device power adjustment command. The actual output power or stored power of the electric energy storage device, thermal storage device, gas internal combustion engine, and electric chiller are collected in real time. This data is collected using sensors installed at the device output or grid connection point, such as current transformers, voltage transformers, power transmitters, and heat meters, with a collection frequency reaching the second level or even higher. The actual power of each device is compared with the power setpoint required in the device power adjustment command corresponding to the current moment to calculate the power deviation value. For example, at 10:32:15, the device power adjustment command for the electric energy storage device requires an output charging power of 10 kW, while the actual collected instantaneous power is 9.2 kW. The power deviation value at this moment is... for Kilowatts. It is understandable that power deviations may occur due to equipment response lag, measurement noise, or unmodeled dynamic characteristics.

[0101] In some embodiments, the power deviation value is input to a proportional-integral-derivative (PID) controller, which outputs a power compensation signal to eliminate the power deviation. The PID controller receives the latest power deviation value and calculates the output at a fixed control cycle, e.g., once per second. In a digital controller, the discretized implementation of the PID control law can be described as follows:

[0102]

[0103] in: This represents the power compensation signal output in the m-th control cycle. This represents the power deviation value obtained from the acquisition and calculation during the m-th control cycle. This represents the sequence of historical power deviation values ​​from the initial time to the m-th control cycle. Indicates the control cycle. , , These are the proportional coefficient, integral coefficient, and derivative coefficient, respectively. The power compensation signal output by the proportional-integral-derivative (PID) controller... It is a continuous adjustment aimed at driving the actual power tracking setpoint. For example, for the aforementioned deviation of -0.8 kW, the proportional-integral-derivative (PID) controller might calculate and output an instantaneous power compensation signal of +0.5 kW, instructing the charging power setpoint of the energy storage device to be temporarily increased by 0.5 kW to offset the negative deviation. Optionally, for devices with high inertia such as gas internal combustion engines and electric refrigeration units, the parameters of the PID controller... , , It can be set to a smaller value to ensure smooth control; for devices with fast power response, such as electric energy storage devices and thermal storage devices, the parameter can be set to a larger value to achieve fast tracking.

[0104] See Figure 4 This is a real-time PID feedback control diagram of the park's energy system, fully demonstrating the dynamic relationship between the command target, actual power acquisition, and PID compensation. From 0 to 20 seconds, the actual power acquisition is slightly lower than the target value (approximately 8-9 kW). The PID compensation signal rapidly increases, continuously outputting positive compensation through a composite control logic of proportional, integral, and derivative, driving the actual power closer to the target value. From 20 to 60 seconds, the increase in the compensation signal and the recovery of the actual power form a positive feedback loop. The compensation signal dynamically fine-tunes itself according to fluctuations in the actual power, ultimately achieving precise tracking of the target power by the actual power. After the compensation signal is superimposed on the baseline command, it drives the actual power to steadily converge to the target value, verifying the effectiveness of the real-time control strategy. It clearly demonstrates how PID control solves problems such as equipment response lag and measurement noise, achieving second-level precise power adjustment.

[0105] In one embodiment of the present invention, for an electric energy storage device, the charging / discharging status command in the device start / stop command, the reference charging / discharging power value in the reference power curve, the power adjustment amount in the device power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command for the electric energy storage device. For a thermal storage device, the heat storage / discharging status command in the device start / stop command, the reference power value in the reference power curve, the power adjustment amount in the device power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command for the thermal storage device. For a gas internal combustion engine and an electric chiller, the start / stop status command in the device start / stop command, the reference operating power value in the reference power curve, the power adjustment amount in the device power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command for the gas internal combustion engine and the electric chiller. The final instantaneous power control command of the electric energy storage device is converted into a drive signal for a power electronic converter to control the charging or discharging process of the electric energy storage battery. The final instantaneous power control command of the thermal storage device is converted into pump and valve opening adjustment signals to control the heat storage or release process of the thermal storage tank. The final instantaneous power control command of the gas internal combustion engine is converted into fuel supply adjustment signals and generator excitation adjustment signals to control the output electrical power of the gas internal combustion engine. The final instantaneous power control command of the electric chiller is converted into compressor frequency adjustment signals to control the cooling power output of the electric chiller.

[0106] In practical implementation, the generation of the final control instruction set and the conversion of instructions into physical drive signals are explained in detail below with examples and data comparisons. In practice, the device start / stop instructions, reference power curve, device power adjustment instructions, and power compensation signals are aggregated to form the final control instruction set for each energy device. This process involves the numerical superposition and logical synthesis of control instructions from various sources. For energy storage devices, the charging / discharging status instructions from the device start / stop instructions, the reference charging / discharging power values ​​from the reference power curve, the power adjustment amount from the device power adjustment instructions, and the power compensation value from the power compensation signals are superimposed to obtain the final instantaneous power control instructions for the energy storage device. The superposition process follows the arithmetic addition principle, but the physical limits of the device must be considered. For example, at a specific moment, the equipment start / stop command obtained from the day-ahead plan requires the energy storage device to be in a "charging" state. The corresponding reference charge / discharge power value in the reference power curve is -50 kW. The equipment power adjustment command obtained from the intraday rolling correction requires an increase of 65 kW. The power compensation signal value obtained from the real-time feedback control is +2 kW. Then, the final instantaneous power control command value of the energy storage device is: (-50) + 65 + 2 = 17 kW. This positive value combined with the "charging" state command means that the energy storage device should be charged at a power of 17 kW. In some embodiments, power superposition is achieved through a command synthesis function, the expression of which is:

[0107]

[0108] in: Indicates at time The final instantaneous power control command value, Indicates at time The reference charge / discharge power value obtained from the reference power curve. Indicates at time The power adjustment amount obtained from the device power adjustment command. Indicates at time The power compensation signal value output from the proportional-integral-derivative (PID) controller. It can be understood that if... If the calculation result is positive and the device start / stop command is "charging", then the final command will be a charging power of [value missing]. ;like If the value is negative and the device start / stop command is "discharge", then the final command will be a discharge power of [value missing]. If the sign of the calculation result conflicts with the state command, the state command shall prevail, and the absolute value of the calculation result shall be taken as the power magnitude, with the sign determined by the state command.

[0109] In some embodiments, for a thermal storage device, the heat storage / release status command in the device start / stop command, the reference power value in the reference power curve, the power adjustment amount in the device power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command of the thermal storage device. The sign convention for the reference power value, power adjustment amount, and power compensation value of the thermal storage device is typically as follows: a positive value represents the power to store heat in the thermal storage device, i.e., the heat storage power; a negative value represents the power to release heat from the thermal storage device, i.e., the heat release power. For example, if the device start / stop command requires "heat storage," the reference power value is 30 kW, the power adjustment amount is -10 kW, and the power compensation value is +0.5 kW, then the final instantaneous power control command value is 30 + (-10) + 0.5 = 20.5 kW. Combined with the "heat storage" status, this means that the thermal storage device should store heat at a power of 20.5 kW. For gas-fired internal combustion engines and electric chillers, the start / stop status command in the equipment start / stop command, the reference operating power value in the reference power curve, the power adjustment amount in the equipment power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command for the gas-fired internal combustion engine and electric chiller. The power values ​​of gas-fired internal combustion engines and electric chillers are usually non-negative, representing output electrical power or cooling power. For example, for a gas-fired internal combustion engine, if the equipment start / stop command is "run," the reference operating power value is 400 kW, the power adjustment amount is +20 kW, and the power compensation value is -1 kW, then the final instantaneous power control command value is 400 + 20 + (-1) = 419 kW, meaning the gas-fired internal combustion engine should output 419 kW of electrical power. It can be understood that after the superposition calculation, the final instantaneous power control command value needs to be compared with the upper and lower limits of equipment operation. If it exceeds the upper limit, it is clamped to the upper limit value; if it is lower than the lower limit, it is clamped to the lower limit value or a shutdown command is triggered.

[0110] In practical implementation, based on the final control command set for each energy device, the physical equipment in the park's integrated energy system is driven to perform corresponding energy production, conversion, storage, and consumption actions. The final instantaneous power control command of the energy storage device is converted into a drive signal for the power electronic converter, controlling the charging or discharging process of the energy storage battery. The conversion process relies on a specific control model of the energy storage device's converter. For example, a 17 kW charging command corresponds to a specific relationship between DC-side voltage and current, achieved by adjusting the duty cycle of the pulse width modulation signal. With command power DC bus voltage There exists a functional relationship, which can be expressed as:

[0111]

[0112] in: This represents the mapping function defined by the converter control logic. It converts the final instantaneous power control command of the thermal storage device into pump and valve opening adjustment signals to control the heat storage or release process of the thermal storage tank. For example, a 20.5 kW thermal storage power command needs to be determined based on the specific heat capacity of the thermal storage medium. ,density and the temperature difference between supply and return water Calculate the required volumetric flow rate of the medium. The calculation formula is:

[0113]

[0114] Where: the volumetric flow rate is calculated Then, the flow-frequency characteristic curve of the water pump is used to convert it into the frequency setpoint of the water pump. And generate the corresponding valve opening command. .

[0115] In practical implementation, the final instantaneous power control command of the gas internal combustion engine is converted into a fuel supply regulation signal and a generator excitation regulation signal to control the output electric power of the gas internal combustion engine. For example, for an output electric power command of 419 kW, the first step is to determine the output power based on the efficiency curve of the gas internal combustion engine. Convert it into the required gas volume flow rate. The conversion formula is:

[0116]

[0117] in: This refers to the lower calorific value of the gas. Gas volumetric flow rate command. The signal is sent to the gas regulating valve as an opening control signal, while the generator excitation system receives the power command. The system maintains stable output voltage and outputs the target electrical power by adjusting the excitation current. The final instantaneous power control command of the electric chiller is converted into a compressor frequency adjustment signal to control the chiller's cooling power output. For example, if an electric chiller receives a final instantaneous power control command of 220 kW for cooling power, the chiller's control system will adjust the power output based on its performance coefficient. This converts it into the input electrical power required by the compressor. The relationship is:

[0118]

[0119] Among them: input power command Then, based on the compressor's power-frequency characteristics, this is converted into the compressor's operating frequency command. Optionally, the compressor's frequency adjustment signal is achieved through a frequency converter, thereby precisely controlling the refrigerant flow and compression ratio to ultimately match the required cooling power output.

[0120] See Figure 5 This diagram illustrates the entire process of command synthesis and its corresponding physical state. Specifically, the blue solid line represents the baseline power value (kW) obtained from the decomposition of the day-ahead optimized scheduling plan. This serves as the basic power reference for the operation of the thermal storage device, reflecting the long-term thermal storage / release strategy based on load and power generation forecasts at the day-ahead scale. The green dashed line represents the power adjustment amount (kW) generated by intraday rolling optimization correction, used to address supply and demand deviations caused by intraday ultra-short-term forecast updates and to finely correct the baseline power. The red dotted line represents the power compensation value (kW) output by the real-time feedback control (PID controller), used to eliminate minute-level deviations between the actual operating power of the equipment and the command setpoint, ensuring control accuracy. The black solid line represents the final instantaneous power control command (kW) obtained by superimposing the above three types of power signals with the equipment start-stop status command. This is the final control quantity driving the physical execution of the thermal storage device. Its calculation follows the arithmetic superposition principle and is generated after being clamped by the upper and lower limits of the equipment's physical operation. The beige-filled area in the figure represents the thermal storage state, which intuitively represents the cumulative thermal storage power and energy storage status of the thermal storage device within 24 hours of the day. It clearly shows the thermal storage process before the morning and evening heating peaks, the energy buffering during the midday low-load period, and the dynamic adjustment effect of the multi-timescale control strategy on the power of the thermal storage device, verifying the effectiveness of the control method in smoothing load fluctuations and optimizing energy allocation.

[0121] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for optimizing the allocation of comprehensive energy supply and demand in industrial parks based on AI multimodal prediction, characterized in that, include: Collect multimodal historical and real-time data of the park's integrated energy system. The multimodal historical and real-time data shall include at least meteorological time series, historical energy load curves, equipment operating condition records, and real-time energy market price series. The collected multimodal historical and real-time data are spatiotemporally aligned and missing values ​​are filled in to form a standardized multimodal data cube; The standardized multimodal data cube is input into a pre-trained AI multimodal prediction model. The AI ​​multimodal prediction model processes different types of data modes in parallel and outputs the predicted values ​​of the combined cooling, heating, and electricity loads of the park for multiple future time periods, as well as the predicted values ​​of the power generation of distributed photovoltaic and wind power. The system receives the load forecast and power generation forecast output by the AI ​​multimodal prediction model, and combines them with the real-time energy market price series to construct a multi-timescale operation optimization model for the park's integrated energy system. Solving the multi-timescale operation optimization model yields an optimized scheduling plan covering multiple time periods, including day-ahead, intraday, and real-time. The optimized scheduling plan includes the charging and discharging power of the electric energy storage device, the heat storage and release power of the thermal storage device, and the output commands of the gas internal combustion engine and the electric chiller.

2. The method for optimizing the allocation of comprehensive energy supply and demand in a park based on AI multimodal prediction as described in claim 1, characterized in that, The spatiotemporal alignment and missing value completion of the collected multimodal historical and real-time data includes: Identify the timestamps and data sampling frequencies of the meteorological time series, historical energy load curves, equipment operating condition records, and real-time energy market price series; Using a preset unified time base and sampling interval, various types of data with different timestamps and sampling frequencies are resampled and interpolated to align all data in the time dimension. Missing data points are detected in the spatiotemporally aligned data sequence. A dual fitting method based on the trend of similar data in adjacent time periods and related modal data is used to generate fill values ​​to fill the missing data points, and finally a standardized multimodal data cube that is continuous in time and correlated in modality is formed.

3. The method for optimizing the allocation of comprehensive energy supply and demand in a park based on AI multimodal prediction as described in claim 2, characterized in that, The standardized multimodal data cube is input into a pre-trained AI multimodal prediction model, including: The AI ​​multimodal prediction model includes meteorological feature extraction branches, load feature extraction branches, equipment operating condition feature extraction branches, and market feature extraction branches, which respectively process the corresponding data in the standardized multimodal data cube; The meteorological feature extraction branch extracts meteorological feature vectors for future periods from meteorological time series; the load feature extraction branch extracts load time series feature vectors from historical energy load curves; the equipment condition feature extraction branch extracts equipment status feature vectors from equipment operating condition records; and the market feature extraction branch extracts price fluctuation feature vectors from real-time energy market price series. The meteorological feature vector, load time series feature vector, equipment status feature vector, and price fluctuation feature vector are fused and input into the spatiotemporal attention prediction network in the AI ​​multimodal prediction model. The spatiotemporal attention prediction network outputs predicted values ​​for the park's cooling load, heating load, electricity load, photovoltaic power generation, and wind power generation for multiple future time periods.

4. The method for optimizing the allocation of comprehensive energy supply and demand in a park based on AI multimodal prediction as described in claim 3, characterized in that, The system receives the load forecast and power generation forecast values ​​output by the AI ​​multimodal prediction model, and combines them with real-time energy market price sequences to construct a multi-timescale operation optimization model for the park's integrated energy system, including: The goal is to minimize the total operating cost of the park's integrated energy system, which includes the cost of purchasing electricity from the upper-level power grid, the cost of natural gas fuel, and the cost of equipment operation and maintenance. The constraints for constructing the multi-timescale operation optimization model include electric power balance constraints, thermal power balance constraints, cold power balance constraints, electric energy storage device operation constraints, thermal storage device operation constraints, and physical operation upper and lower limits and ramp rate constraints for key energy conversion equipment such as gas internal combustion engines and electric chillers. The multi-timescale operation optimization model divides the next day into three timescales: day-ahead, intraday, and real-time. At the day-ahead timescale, it makes decisions on equipment start-up and shutdown and energy transfer over long time scales. At the intraday timescale, it performs rolling optimization of power plans over several hours. At the real-time timescale, it handles power deviation balancing at the minute level.

5. The method for optimizing the allocation of comprehensive energy supply and demand in a park based on AI multimodal prediction as described in claim 4, characterized in that, Solving the multi-timescale operation optimization model yields an optimized scheduling plan covering multiple time periods, including: The decomposition and coordination algorithm is used to solve the multi-timescale operation optimization model, decomposing the original problem into a daily main problem and multiple intraday sub-problems; Solving the main problem of the day, we obtain the planned charging and discharging power curves of the electric energy storage device, the planned heat storage and release power curves of the thermal storage device, and the planned output curves of the gas internal combustion engine and the electric chiller for the next 24 hours at one-hour intervals. Based on the solution results of the main problem of the day, the intraday sub-problems covering the next few hours are solved in a rolling manner, and the planned charging and discharging power curves of the energy storage equipment, the planned heat storage and release power curves of the thermal storage device, and the planned output curves of the key equipment are refined and corrected to form an intraday rolling plan with a 15-minute interval. The solution to the day-ahead master problem and the intraday rolling plan together constitute an optimized scheduling plan.

6. The method for optimizing the allocation of comprehensive energy supply and demand in a park based on AI multimodal prediction as described in claim 5, characterized in that, Also includes: The daytime schedule in the optimized scheduling plan is decomposed into specific equipment start-up and shutdown instructions and baseline power curves, and then sent to each energy conversion and storage device for execution. During the intraday execution phase, based on the latest ultra-short-term load forecast and power generation forecast, the intraday rolling plan in the optimized scheduling plan is revised to generate equipment power adjustment instructions within the current time window; During the real-time execution phase, the actual operating power of each energy conversion and storage device is monitored and compared with the device power adjustment command. Power compensation signals are generated through real-time feedback control. The device start / stop commands, reference power curves, device power adjustment commands, and power compensation signals are aggregated to form the final control command set for each energy device; Based on the final set of control instructions for each energy device, the physical devices in the park's integrated energy system are driven to perform corresponding energy production, conversion, storage, and consumption actions, thereby completing the optimal allocation of supply and demand. The step of decomposing the day-ahead plan in the optimized scheduling plan into specific equipment start / stop instructions and baseline power curves includes: The planned charge and discharge power curves of the energy storage devices in the day-ahead portion of the optimized scheduling plan are analyzed to generate charge and discharge status commands and baseline charge and discharge power values ​​for the energy storage devices at the start of each hour. The thermal storage and release power curves of the thermal storage devices in the day-ahead portion of the optimized scheduling plan are analyzed to generate thermal storage or release status commands and reference power values ​​for the thermal storage devices at the start of each hour. The planned output curves of the gas internal combustion engine and electric chiller in the daytime portion of the optimized scheduling plan are analyzed to generate start-stop status commands and baseline operating power values ​​for the gas internal combustion engine and electric chiller within each hour.

7. The method for optimizing the allocation of comprehensive energy supply and demand in a park based on AI multimodal prediction as described in claim 6, characterized in that, During the intraday execution phase, based on the latest ultra-short-term load and power generation forecasts, the intraday rolling schedule in the optimized dispatch plan is revised, including: During the intraday execution phase, the AI ​​multimodal prediction model is invoked to perform ultra-short-term predictions, obtaining updated cold, heat, and electricity load forecasts and photovoltaic and wind power generation forecasts for the next few hours. The relevant parameters in the multi-timescale operation optimization model are updated with the latest ultra-short-term forecast values, and the intraday rolling plan for the next few hours is re-solved to obtain the updated equipment power plan curve. Compare the updated equipment power plan curve with the intraday rolling plan in the original optimized scheduling plan, and calculate the power adjustment amount for each equipment; Based on the power adjustment amount of each device, generate equipment power adjustment instructions for electric energy storage devices, thermal storage devices, gas internal combustion engines, and electric chillers within the current time window.

8. The method for optimizing the allocation of comprehensive energy supply and demand in a park based on AI multimodal prediction as described in claim 7, characterized in that, During the real-time execution phase, the actual operating power of each energy conversion and storage device is monitored and compared with the device power adjustment command. A power compensation signal is generated through real-time feedback control, including: Real-time acquisition of the actual output power or storage power of electric energy storage devices, thermal storage devices, gas internal combustion engines and electric chillers; The actual power of each device is compared with the power setting value required in the device power adjustment command at the current time, and the power deviation value is calculated. The power deviation value is input to a proportional-integral-derivative (PID) controller, which outputs a power compensation signal to eliminate the power deviation.

9. The method for optimizing the allocation of comprehensive energy supply and demand in a park based on AI multimodal prediction as described in claim 8, characterized in that, The device start / stop commands, reference power curves, device power adjustment commands, and power compensation signals are aggregated to form the final control command set for each energy device, including: For an energy storage device, the charging and discharging status command in the device start / stop command, the reference charging and discharging power value in the reference power curve, the power adjustment amount in the device power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command of the energy storage device. For a thermal storage device, the thermal storage and release status command in the device start / stop command, the reference power value in the reference power curve, the power adjustment amount in the device power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command of the thermal storage device. For both gas internal combustion engines and electric chillers, the start / stop status command in the equipment start / stop command, the reference operating power value in the reference power curve, the power adjustment amount in the equipment power adjustment command, and the power compensation value in the power compensation signal are superimposed to obtain the final instantaneous power control command for the gas internal combustion engine and the electric chiller.

10. The method for optimizing the allocation of comprehensive energy supply and demand in a park based on AI multimodal prediction as described in claim 9, characterized in that, Based on the final set of control instructions for each energy device, the physical devices in the park's integrated energy system are driven to perform corresponding energy production, conversion, storage, and consumption actions, thereby achieving optimized supply and demand configuration, including: The final instantaneous power control command of the energy storage device is converted into a drive signal of the power electronic converter to control the charging or discharging process of the energy storage battery. The final instantaneous power control command of the heat storage device is converted into water pump and valve opening adjustment signals to control the heat storage or heat release process of the heat storage tank. The final instantaneous power control command of the gas internal combustion engine is converted into a fuel supply regulation signal and a generator excitation regulation signal to control the output electric power of the gas internal combustion engine. The final instantaneous power control command of the electric chiller is converted into a compressor frequency adjustment signal to control the cooling power output of the electric chiller.

Images