An artificial intelligence-based photovoltaic storage direct supply data center prediction control method

Abstract

The application discloses an artificial intelligence-based photovoltaic storage direct supply data center prediction control method, relates to the technical field of intelligent control, and comprises the following steps: inputting a multi-physical field coupling input feature set into a POD dimension reduction process of physical information constraints, extracting dominant modes under the constraints of energy conservation, convection heat exchange and boundary conditions, and obtaining low-dimensional POD coefficients through mode projection; inputting the low-dimensional POD coefficients into a multi-time scale prediction model set to obtain POD coefficient prediction values of multiple future time periods; and reconstructing load power sequences, photovoltaic output sequences and energy storage SOC sequences according to the POD coefficient prediction values of the multiple future time periods and a physical information POD mode library. The application improves the adaptability of real-time control instructions and the continuity of the prediction control process.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a predictive control method for direct optical storage data center based on artificial intelligence. Background Technology

[0002] As the synergistic application of photovoltaic power generation, energy storage regulation, DC power distribution and immersion liquid cooling technology in data center scenarios continues to deepen, the related control methods have gradually evolved from energy management based on fixed rules to intelligent predictive control paths that integrate multi-source monitoring, load forecasting and rolling scheduling, and have begun to focus on the coupling relationship between electric power field, thermal field, flow field and environmental field.

[0003] The main problem with existing related technologies is that high-dimensional sensor data is mostly modeled directly using pure data-driven models, which lacks constraints on energy conservation, convective heat transfer and boundary conditions. This makes it difficult to stably compress coupling characteristics and maintain physical consistency, resulting in a decrease in the prediction accuracy and control robustness of pure data-driven models under new operating conditions, extreme weather or sudden load changes. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based predictive control method for direct-supply optical-storage data centers to solve the problem of achieving physically consistent predictive control of high-dimensional coupled multi-physics data in direct-supply optical-storage data centers.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an artificial intelligence-based predictive control method for direct photovoltaic-storage data center supply. The method includes: collecting photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data, power supply matching feedback data, weather forecast data, and immersion liquid cooling operation data; performing noise reduction, alignment, and normalization processing; constructing a multi-physics field input feature set based on the coordinated changes of the electric power field, thermal field, flow field, and environmental field; inputting the multi-physics field coupled input feature set into a physical information-constrained POD dimensionality reduction process; extracting the dominant modes under energy conservation constraints, convective heat transfer constraints, and boundary condition constraints; and obtaining low-dimensional POD coefficients through mode projection; and then processing the low-dimensional POD coefficients... Input a multi-timescale prediction model set to obtain POD coefficient predictions for multiple future time periods; based on the predicted POD coefficients for multiple future time periods and the physical information POD modality library, reconstruct the load power sequence, photovoltaic output sequence, and energy storage SOC sequence; construct a linear programming optimization model using the load power sequence, photovoltaic output sequence, and energy storage SOC sequence as inputs to obtain energy storage charging and discharging commands and grid power extraction commands; convert the energy storage charging and discharging commands and grid power extraction commands into real-time control commands and execute them; after execution, collect the actual operating results, and perform correction and update processing on the physical information POD modality library and the multi-timescale prediction model set based on the actual operating results, load power sequence, photovoltaic output sequence, and energy storage SOC sequence.

[0007] As a preferred embodiment of the AI-based optical-storage direct-supply data center predictive control method of the present invention, wherein: the construction of a multi-physics coupled input feature set specifically includes, Collect photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data, power supply matching feedback data, weather forecast data, and immersion liquid cooling operation data to form multi-source operation monitoring data; Multi-source operation monitoring data are uniformly aggregated and then processed by denoising, alignment and normalization to form a multi-physics field coupled input feature set; The multiphysics field includes the electric power field, thermal field, flow field, and environmental field.

[0008] As a preferred embodiment of the AI-based optical-storage direct-supply data center predictive control method of the present invention, wherein obtaining the low-dimensional POD coefficient specifically includes, Based on the feature data representing power change, heat generation change, coolant flow, heat transfer state, chip surface heat flux density and coolant inlet temperature in the multi-physics field coupled input feature set, energy conservation constraints, convective heat transfer constraints and boundary condition constraints are constructed to generate a set of physical constraint conditions. Based on the set of physical constraints, singular value decomposition is performed on the multi-physics field coupled input feature set to extract the dominant modes that satisfy the preset cumulative energy ratio and generate a dominant mode set. A physical information POD mode library is constructed based on the dominant mode set, and the multi-physics field coupled input feature set is projected onto the mode space corresponding to the physical information POD mode library to obtain low-dimensional POD coefficients.

[0009] As a preferred embodiment of the AI-based optical storage direct-supply data center predictive control method of the present invention, wherein obtaining the predicted POD coefficient values ​​for multiple future time periods specifically includes, According to the preset prediction task division rules, the prediction process of the low-dimensional POD coefficient is divided to generate a multi-timescale prediction process set corresponding to rapid response control, rolling coordinated adjustment and operation trend prediction respectively. Based on the multi-timescale prediction process set, and using the shared physical information POD mode basis function as a foundation, GA-BPNN prediction structures are constructed for each prediction process to generate a multi-timescale prediction model set. By inputting the low-dimensional POD coefficients into a multi-timescale prediction model set, predicted POD coefficient values ​​for multiple future time periods can be obtained.

[0010] As a preferred embodiment of the AI-based optical-storage direct-supply data center predictive control method of the present invention, wherein: the generation of a multi-time-scale predictive model set specifically includes, Extract the prediction objects and parameter update requirements corresponding to each prediction process from the multi-timescale prediction process set, and generate a task configuration subset; Based on the task configuration subset and using the shared physical information POD modal basis function as input, the network layer, node configuration, genetic algorithm optimization range and parameter update rules corresponding to each prediction process are configured to generate the prediction structure configuration result. Based on the prediction structure configuration results, a corresponding GA-BPNN prediction structure is constructed for each prediction process, and a physical driving branch is configured in each GA-BPNN prediction structure. By combining each GA-BPNN prediction structure with the physics-driven branch, a set of multi-timescale prediction models is generated.

[0011] As a preferred embodiment of the AI-based photovoltaic-storage direct-supply data center predictive control method of the present invention, wherein the reconstructing of the load power sequence, photovoltaic output sequence, and energy storage SOC sequence specifically includes, Call the first K orders of physical information POD modes from the physical information POD mode library; The predicted POD coefficients for multiple future time periods are reconstructed by tensor product with the POD modes of the first K physical information to generate multi-physics collaborative reconstruction results. The corresponding reconstruction components in the multi-physics collaborative reconstruction results are classified and mapped to obtain load power sequence, photovoltaic power output sequence and energy storage SOC sequence.

[0012] As a preferred embodiment of the AI-based optical-storage direct-supply data center predictive control method of the present invention, wherein: the process of obtaining energy storage charging and discharging commands and mains power supply commands specifically includes, Using load power sequence, photovoltaic output sequence and energy storage SOC sequence as inputs, a linear programming solution is constructed according to the optimization objectives of reducing grid electricity consumption and reducing curtailment of solar power. Based on the linear programming solution, under the constraints of power balance, energy storage SOC dynamic constraints, charging and discharging power limits, energy storage SOC safe range constraints, and mains power upper limit constraints, the energy storage charging power, energy storage discharging power, and mains power extraction power for each time period are solved to generate time period optimization results. Extract the energy storage charging / discharging command and mains power draw command corresponding to the current time from the time period optimization results.

[0013] As a preferred embodiment of the AI-based optical-storage direct-supply data center predictive control method of the present invention, the step of converting energy storage charging and discharging commands and mains power extraction commands into real-time control commands and executing them specifically includes: Configure the corresponding control content of the bidirectional DC / DC converter according to the energy storage charging and discharging command, and configure the corresponding start / stop content and power setting content of the bidirectional AC / DC converter according to the mains power supply command to generate a real-time control command set; The real-time control command set is sent to the corresponding execution object for control execution. After the control execution, the photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data and power supply matching feedback data in the multi-source operation monitoring data are collected back to obtain the actual operation results.

[0014] As a preferred embodiment of the AI-based optical storage direct-supply data center predictive control method of the present invention, the step of performing correction and update processing on the Physical Information Device (POD) modality library and the multi-timescale prediction model set specifically includes: Based on the actual operating results and the load power sequence, photovoltaic output sequence and energy storage SOC sequence, the status changes in power supply matching feedback data, data center load data and multi-source operation monitoring data are correlated and analyzed to generate status identification input results. Based on the status recognition input results, the power supply deviation status and operating condition change status are identified, and status recognition results are generated. Based on the state identification results, short-term error correction processing is performed on the power supply deviation state, and medium-term model update processing is performed on the operating condition change state when the preset daily update cycle is reached. When the preset long-term update cycle is reached, long-term adaptive evolution processing is performed on the operating condition change state to generate correction update processing results.

[0015] As a preferred embodiment of the AI-based optical storage direct-supply data center predictive control method of the present invention, wherein: the generation, correction, and update processing results are used to update the Physical Information Device (POD) modality library and the multi-timescale prediction model set, specifically including: Based on the correction and update processing results, the newly added feedback data from the actual operation results and the seasonal characteristics, photovoltaic module degradation characteristics, energy storage battery aging characteristics and IT configuration replacement characteristics from the multi-source operation monitoring data are collected to generate updated input results. Based on the updated input results, the POD dimensionality reduction process with physical information constraints is re-executed when the preset daily update cycle is reached, generating modal update results; Based on the modal update results, the multi-timescale prediction model set is retrained when the preset daily update cycle is reached. When the preset long-term update cycle is reached, seasonal features, photovoltaic module degradation features, energy storage battery aging features, and IT configuration replacement features are introduced to fully retrain the multi-timescale prediction model set, resulting in the updated physical information POD modal library and multi-timescale prediction model set.

[0016] The beneficial effects of this invention are as follows: By inputting the multi-physics field coupled input feature set into the physical information constraint POD dimensionality reduction process, the compressed representation of multi-physics field coupled information is realized, which is used to generate low-dimensional POD coefficients and physical information POD mode library, thereby providing a stable foundation for multi-timescale prediction model set and linear programming solution formula; and by performing correction and update processing on physical information POD mode library and multi-timescale prediction model set, the closed-loop connection between actual operation results and predictive control is realized, which is used to correct power supply deviation state and operating condition change state, thereby improving the adaptability of real-time control commands and the continuity of predictive control process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a predictive control method for direct optical storage power supply to data centers based on artificial intelligence.

[0019] Figure 2A flowchart for constructing a multiphysics coupling input feature set.

[0020] Figure 3 The flowchart for obtaining low-dimensional POD coefficients.

[0021] Figure 4 A flowchart for generating predicted POD coefficients and reconstructing sequences for multiple future time periods. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an artificial intelligence-based predictive control method for direct optical storage power to data centers, comprising the following steps:

[0026] S1. Collect photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data, power supply matching feedback data, weather forecast data, and immersion liquid cooling operation data, and perform noise reduction, alignment, and normalization processing. Based on the coordinated change relationship of electric power field, thermal field, flow field, and environmental field, construct a multi-physics field input feature set.

[0027] S1.1 Collect photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data, power supply matching feedback data, weather forecast data, and immersion liquid cooling operation data to form multi-source operation monitoring data.

[0028] Specifically, photovoltaic array operation data is collected from the power acquisition terminal corresponding to the photovoltaic array, energy storage power station status data is collected from the status acquisition terminal corresponding to the energy storage power station, DC bus status data is collected from the voltage and current acquisition terminal corresponding to the DC bus, mains power standby status data is collected from the mains power access side, data center load data is collected from the power monitoring interfaces corresponding to the rack-level smart PDU, the branch energy meter of the rack-head cabinet, and the servers, switches, and storage devices, power supply matching feedback data is collected from the power supply status verification location, weather forecast data is collected from the weather forecast release location, and immersion liquid cooling operation data is collected from the temperature and flow acquisition terminal corresponding to the immersion liquid cooling. Then, the photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data, power supply matching feedback data, weather forecast data, and immersion liquid cooling operation data are time-correlated and field-merged to form multi-source operation monitoring data.

[0029] S1.2. The multi-source operation monitoring data are uniformly aggregated and processed by denoising, alignment and normalization to form a multi-physics field coupled input feature set.

[0030] Specifically, the multi-source operation monitoring data is aggregated and organized according to a unified field order and a unified time benchmark. First, missing segments, abnormal jump segments, and duplicate segments are removed to obtain denoised multi-source operation monitoring data. Then, the denoised multi-source operation monitoring data is time-series matched and sampled according to the unified time benchmark to obtain aligned multi-source operation monitoring data. The aligned multi-source operation monitoring data is then subjected to dimensional conversion and numerical normalization according to power quantity, temperature quantity, flow quantity, state quantity, and feedback quantity to obtain a multi-physics field coupled input feature set that can jointly characterize the energy supply process, load process, and heat exchange process. The multi-physics field includes electric power field, thermal field, flow field, and environmental field.

[0031] S2. The POD dimensionality reduction process involves inputting the multi-physics field coupled input feature set into the physical information constraint. Under the constraints of energy conservation, convection heat transfer, and boundary conditions, the dominant modes are extracted, and the low-dimensional POD coefficients are obtained through mode projection.

[0032] S2.1 Based on the characteristic data representing power change, heat generation change, coolant flow, heat transfer state, chip surface heat flux density and coolant inlet temperature in the multi-physics field coupled input feature set, construct energy conservation constraints, convective heat transfer constraints and boundary condition constraints, and generate a set of physical constraint conditions.

[0033] Specifically, based on the feature data representing power changes, heat generation changes, coolant flow, heat transfer state, chip surface heat flux density, and coolant inlet temperature from the multi-physics field coupled input feature set, the correspondence between power input, power output, and heat generation is extracted. Energy conservation constraints are generated according to the balance relationship between total energy supply, total energy consumption, and total heat generation. The flow rate change corresponding to coolant flow, the temperature change corresponding to heat transfer state, and the heat transfer relationship are extracted. Convective heat transfer constraints are constructed based on the coolant flow rate, temperature change, and heat transfer relationship. Boundary condition constraints are constructed based on the boundary range and change relationship of chip surface heat flux density and coolant inlet temperature. Finally, the energy conservation constraints, convective heat transfer constraints, and boundary condition constraints are unified and aggregated to generate a set of physical constraint conditions.

[0034] It should also be noted that the boundary range refers to the allowable range of variation in the chip surface heat flux density and coolant inlet temperature during predictive control.

[0035] S2.2. Based on the set of physical constraints, perform singular value decomposition on the multi-physics field coupled input feature set, extract the dominant modes that satisfy the preset cumulative energy ratio, and generate the dominant mode set.

[0036] Specifically, based on the set of physical constraints, each characteristic component in the multi-physics coupling input feature set is matched with the energy conservation constraint, convection heat transfer constraint, and boundary condition constraint. First, the multi-physics coupling input feature set that satisfies the set of physical constraints is matrixed to obtain the eigenvalue matrix. Then, singular value decomposition is performed on the eigenvalue matrix to obtain the candidate mode set arranged according to the size of energy contribution and the corresponding singular value sequence. Based on the singular value sequence, the single-mode energy contribution ratio and the cumulative energy contribution ratio corresponding to each candidate mode are calculated. The cumulative energy contribution ratio is then compared with the preset cumulative energy ratio item by item. The candidate mode set that reaches the preset cumulative energy ratio is retained to generate the dominant mode set.

[0037] It should also be noted that the preset cumulative energy percentage refers to the cumulative energy contribution ratio threshold set in advance to ensure that the dominant mode set can characterize the main change characteristics of the multi-physics coupled input feature set;

[0038] Based on the set of physical constraints, a constrained singular value decomposition and dominant mode screening are performed on the multiphysics coupled input feature set to form a physical information-constrained POD dimensionality reduction process. The physical information-constrained POD dimensionality reduction process can be expressed as follows:

[0039] ;

[0040] in, This represents the overall optimization objective value corresponding to the POD dimensionality reduction process with physical information constraints. This indicates that the goal is to minimize the overall optimization objective value. This represents the feature matrix composed of multi-physics coupled input feature sets. Indicates from the previous The mode matrix composed of the dominant modes of the first order. This represents the low-dimensional POD coefficient matrix. Describing the F-norm, Indicates the physical constraint weight coefficient. This represents the physical constraint term, which consists of energy conservation constraints, convective heat transfer constraints, and boundary condition constraints, and the physical constraint term acts on the mode matrix. ;

[0041] The cumulative energy contribution percentage for each candidate mode can be expressed as:

[0042] ;

[0043] in, Indicates the preceding The cumulative energy contribution percentage corresponding to the candidate modes of the first order. Indicates the first Singular values ​​corresponding to the candidate modes of order. Indicates the total number of candidate modes. This indicates the order of the candidate modes involved in the cumulative calculation. Indicates the order number of the candidate mode.

[0044] S2.3 Construct a Physical Information POD Mode Library based on the dominant mode set, and project the multi-physics field coupled input feature set onto the mode space corresponding to the Physical Information POD Mode Library to obtain low-dimensional POD coefficients.

[0045] Specifically, based on the dominant mode set, each dominant mode is labeled with a mode number according to the dominant mode retention order. The modal components, single-mode energy contribution ratio, and cumulative energy contribution ratio corresponding to each dominant mode are extracted and stored in a unified field order to establish a physical information POD mode library. According to the mode arrangement order in the physical information POD mode library, the multiphysics coupling input feature set is mapped to the corresponding mode space in the physical information POD mode library. The modal projection method is used to calculate the projection components of the multiphysics coupling input feature set in each dominant mode direction, and the projection components are combined according to the mode number to obtain low-dimensional POD coefficients.

[0046] S3. Input the low-dimensional POD coefficients into the multi-timescale prediction model set to obtain the predicted POD coefficient values ​​for multiple future time periods.

[0047] S3.1. Based on the preset prediction task division rules, the prediction process of the low-dimensional POD coefficient is divided to generate a multi-timescale prediction process set corresponding to rapid response control, rolling collaborative adjustment and operation trend prediction.

[0048] Specifically, based on the preset prediction task division rules, the task division criteria are extracted from the control response time, adjustment update interval, and prediction coverage duration corresponding to the low-dimensional POD coefficients. The prediction tasks of the low-dimensional POD coefficients are then hierarchically classified in ascending order of control response time, adjustment update interval, and prediction coverage duration. The prediction task with the shortest control response time and used for rapid generation of the current control command is identified as the prediction process corresponding to fast response control. The prediction task with a medium adjustment update interval and used for continuous adjustment and correction is identified as the prediction process corresponding to rolling coordinated adjustment. The prediction task with the longest prediction coverage duration and used for advance judgment of subsequent operating condition changes is identified as the prediction process corresponding to operation trend prediction. The prediction objectives, prediction time domains, and update requirements of the prediction processes corresponding to fast response control, rolling coordinated adjustment, and operation trend prediction are extracted respectively. The processes are then merged according to the one-to-one correspondence between process category and prediction objectives, prediction time domains, and update requirements to generate a multi-timescale prediction process set.

[0049] It should also be noted that the preset forecast task division rules refer to the pre-set basis for dividing the forecast process for different forecast needs corresponding to rapid response control, rolling collaborative adjustment and operation trend prediction.

[0050] S3.2 Based on the multi-timescale prediction process set, and using the shared physical information POD modal basis function as a foundation, construct GA-BPNN prediction structures for each prediction process to generate a multi-timescale prediction model set.

[0051] S3.2.1 Extract the prediction objects and parameter update requirements corresponding to each prediction process from the multi-timescale prediction process set, and generate a task configuration subset.

[0052] Specifically, the prediction processes corresponding to rapid response control, rolling coordinated adjustment, and operation trend prediction are read from the multi-timescale prediction process set. The prediction objects, prediction time domains, and parameter update frequencies of each prediction process are extracted. The prediction processes are classified, organized, and merged according to the correspondence between the prediction objects and parameter update requirements, so that the prediction objects and parameter update requirements of each prediction process form a one-to-one configuration relationship, and a task configuration subset is generated.

[0053] S3.2.2 Based on the task configuration subset and using the shared physical information POD modal basis function as input, configure the network level, node configuration, genetic algorithm optimization range and parameter update rules corresponding to each prediction process to generate the prediction structure configuration result.

[0054] Specifically, based on the prediction objects and parameter update requirements corresponding to each prediction process in the task configuration subset, the input dimensions and modal arrangement relationships corresponding to the shared physical information POD modal basis functions are used as the configuration benchmark. The network layers corresponding to each prediction process are allocated, the node configurations corresponding to each prediction process are matched, the optimization range of the genetic algorithm corresponding to each prediction process is set, and the parameter update rules corresponding to each prediction process are organized accordingly. The network layers, node configurations, genetic algorithm optimization ranges, and parameter update rules are merged item by item according to the prediction process to generate the prediction structure configuration results.

[0055] It should also be noted that the Physical Information POD modal basis functions are a set of modal basis vectors composed of the modal components corresponding to each dominant mode in the Physical Information POD modality library. These vectors are used to uniformly determine the input dimension and modal arrangement order of each prediction process and serve as the common input basis for each GA-BPNN prediction structure when receiving low-dimensional POD coefficients. The network layers refer to the input layer, hidden layer, and output layer in each GA-BPNN prediction structure. The input layer is used to receive low-dimensional POD coefficients, the output layer is used to output the prediction results of the corresponding prediction object, and the hidden layer is used to establish the nonlinear mapping relationship between the input and output. The hidden layer can be set as a single hidden layer or multiple hidden layers. The optimization range of the genetic algorithm refers to the range of values ​​for searching the initial weights, bias parameters, and connection parameters in each GA-BPNN prediction structure. This range is determined based on the number of nodes in the input layer, the number of nodes in the hidden layer, the number of nodes in the output layer, and the historical training convergence results.

[0056] S3.2.3 Based on the prediction structure configuration results, construct the corresponding GA-BPNN prediction structure for each prediction process, and configure the physical driving branch in each GA-BPNN prediction structure.

[0057] Specifically, based on the network level, node configuration, genetic algorithm optimization range, and parameter update rules corresponding to each prediction process in the prediction structure configuration results, GA-BPNN prediction structures matching the prediction processes corresponding to rapid response control, rolling collaborative adjustment, and operation trend prediction are constructed respectively. Each GA-BPNN prediction structure is optimized and configured according to the genetic algorithm optimization range corresponding to each prediction process in the prediction structure configuration results. The update method of each GA-BPNN prediction structure is set accordingly based on the parameter update rules corresponding to each prediction process in the prediction structure configuration results. Physical driving branches are configured in each GA-BPNN prediction structure to characterize the correlation between power change, heat generation change, coolant flow, heat transfer state, chip surface heat flux density, and coolant inlet temperature, so that each GA-BPNN prediction structure and the physical driving branches form a combination relationship corresponding to the prediction process.

[0058] S3.2.4 Combine each GA-BPNN prediction structure with the physical driving branch to generate a multi-timescale prediction model set.

[0059] Specifically, the corresponding GA-BPNN prediction structures and physical driving branches are paired item by item according to the prediction process correspondence. Based on the one-to-one correspondence of prediction process category, prediction object, prediction time domain, and parameter update requirements, the GA-BPNN prediction structure corresponding to fast response control is combined with the physical driving branch corresponding to fast response control, the GA-BPNN prediction structure corresponding to rolling collaborative adjustment is combined with the physical driving branch corresponding to rolling collaborative adjustment, and the GA-BPNN prediction structure corresponding to operation trend prediction is combined with the physical driving branch corresponding to operation trend prediction. The combined results corresponding to each prediction process are then uniformly merged and sorted according to the process order in the multi-timescale prediction process set, so that the combined results corresponding to each prediction process form a whole model set that can be jointly executed for prediction, generating a multi-timescale prediction model set.

[0060] S3.3 Input the low-dimensional POD coefficient into the multi-timescale prediction model set to obtain the predicted POD coefficient values ​​for multiple future time periods.

[0061] Specifically, the low-dimensional POD coefficients are input into a multi-timescale prediction model set in chronological order. The prediction structure corresponding to rapid response control uses a single-step forward prediction method based on the low-dimensional POD coefficients at the current moment to calculate the predicted POD coefficient value for the next control moment. The prediction structure corresponding to rolling coordinated regulation uses a rolling forward prediction method based on a sliding time window to continuously calculate the predicted POD coefficient values ​​for multiple adjacent time periods. The prediction structure corresponding to operational trend prediction uses a multi-step trend extrapolation prediction method based on a longer historical time series input to calculate the predicted POD coefficient values ​​for a longer prediction time domain. The predicted POD coefficient values ​​output by each prediction structure are then spliced ​​and arranged in chronological order to obtain the predicted POD coefficient values ​​for multiple future time periods.

[0062] It should also be noted that the single-step forward prediction method, the rolling forward prediction method, and the multi-step trend extrapolation prediction method are all calculated by combining the corresponding GA-BPNN prediction structure with the physical driving branch.

[0063] S4. Based on the predicted POD coefficient values ​​for multiple future time periods and the physical information POD mode library, reconstruct the load power sequence, photovoltaic output sequence, and energy storage SOC sequence.

[0064] S4.1 Call the first K orders of physical information POD modes from the physical information POD mode library.

[0065] Specifically, according to the mode number and mode arrangement order in the Physical Information POD mode library, the Physical Information POD modes of each order in the Physical Information POD mode library are read sequentially. The first K order Physical Information POD modes are filtered according to the corresponding number range. The filtered first K order Physical Information POD modes are organized and arranged according to the mode number to ensure that the first K order Physical Information POD modes are consistent with the mode calling order required for subsequent tensor product reconstruction, thus completing the calling of the first K order Physical Information POD modes in the Physical Information POD mode library.

[0066] It should also be noted that K is the minimum modal order that makes the cumulative energy contribution ratio of the first K physical information POD modes reach the preset cumulative energy contribution ratio threshold. The value is determined based on the magnitude of the singular values ​​and the cumulative energy contribution ratio of each dominant mode.

[0067] S4.2. The predicted POD coefficient values ​​for multiple future time periods are reconstructed by tensor product with the POD modes of the first K physical information to generate multi-physics field collaborative reconstruction results. The corresponding reconstruction components in the multi-physics field collaborative reconstruction results are classified and mapped to obtain the load power sequence, photovoltaic power output sequence and energy storage SOC sequence.

[0068] Specifically, the predicted POD coefficient values ​​for multiple future time periods are mapped one by one to the first K physical information POD modes according to the prediction time period order. Tensor products are calculated for the coefficient values ​​of each time period in the predicted POD coefficient values ​​for multiple future time periods and the modal components of each order in the first K physical information POD modes to obtain the modal reconstruction components corresponding to each prediction time period. The modal reconstruction components corresponding to each prediction time period are superimposed according to the modal order to obtain the time period reconstruction results corresponding to the prediction time period. The reconstruction results of each time period are arranged and organized according to the time period order to generate multi-physics collaborative reconstruction results. Based on the correspondence between each reconstruction component in the multi-physics collaborative reconstruction results and the original feature dimensions, the load corresponding component is extracted from the reconstruction component representing the data center load change and merged in time order to obtain the load power sequence; the photovoltaic corresponding component is extracted from the reconstruction component representing the photovoltaic output change and merged in time order to obtain the photovoltaic power output sequence; the energy storage corresponding component is extracted from the reconstruction component representing the energy storage state change and merged in time order to obtain the energy storage SOC sequence.

[0069] S5. Construct a linear programming optimization model with the load power sequence, photovoltaic output sequence and energy storage SOC sequence as inputs to obtain the energy storage charging and discharging commands and the mains power supply commands.

[0070] S5.1. Using the load power sequence, photovoltaic output sequence, and energy storage SOC sequence as inputs, construct a linear programming solution according to the optimization objectives of reducing grid power consumption and reducing curtailment of solar power.

[0071] Specifically, using the load power sequence, photovoltaic output sequence, and energy storage SOC sequence as unified inputs, the system reads the load demand for each time period in the load power sequence, the photovoltaic supply for each time period in the photovoltaic output sequence, and the energy storage status for each time period in the energy storage SOC sequence in chronological order. Then, the load demand, photovoltaic supply, and energy storage status for each time period are matched and combined to obtain a set of time-period supply and demand relationships for solving the problem. Focusing on the optimization goals of reducing grid electricity consumption and reducing curtailment of photovoltaic power, the system linearizes the power distribution relationship, energy storage charging and discharging relationship, and grid power supplementation relationship for each time period in the set of time-period supply and demand relationships. Finally, the linearized relationship of variables for each time period is merged according to a unified solution format to generate a linear programming solution.

[0072] S5.2 Based on the linear programming solution, under the constraints of power balance, energy storage SOC dynamic constraints, charging and discharging power limits, energy storage SOC safe range constraints, and mains power upper limit constraints, the energy storage charging power, energy storage discharging power, and mains power extraction power for each time period are solved to generate time period optimization results.

[0073] Specifically, the load power sequence, photovoltaic output sequence, and energy storage SOC sequence for each time period are read in chronological order. These sequences are then substituted into a linear programming solution to form the power allocation relationship for each time period. Based on this power allocation relationship, and considering power balance constraints, energy storage SOC dynamic constraints, charging and discharging power limits, energy storage SOC safety range constraints, and grid power upper limit constraints, the solvable constraints for each time period are determined. Based on these solvable constraints, the energy storage charging power, energy storage discharging power, and grid power consumption for each time period are calculated. Finally, the energy storage charging power, energy storage discharging power, and grid power consumption for each time period are arranged in chronological order to generate the time period optimization results.

[0074] It should also be noted that the power balance constraint can be expressed as:

[0075] ;

[0076] in, Indicates time Photovoltaic power output, Indicates time The power output from the mains. Indicates time The energy storage discharge power, Indicates time Energy storage charging power, Indicates time Load power, An index representing a time point;

[0077] The dynamic constraints of energy storage SOC can be expressed as:

[0078] ;

[0079] in, This indicates the State of Charge (SOC) of the energy storage at the next moment. Indicates time Energy storage SOC, Indicates energy storage charging efficiency. Indicates the time interval between adjacent time periods. Indicates the rated capacity of energy storage. Indicates the energy storage discharge efficiency;

[0080] The charge / discharge power limitation constraint can be expressed as:

[0081] ;

[0082] ;

[0083] in, Indicates the maximum charging power of the energy storage. Indicates the maximum discharge power of the energy storage;

[0084] The safety range constraint of energy storage SOC can be expressed as:

[0085] ;

[0086] in, Indicates the lower limit of the energy storage SOC. Indicates the upper limit of the energy storage SOC;

[0087] The upper limit constraint on mains power can be expressed as:

[0088] ;

[0089] in, This indicates the upper limit of the mains power output.

[0090] S5.3 Extract the energy storage charging / discharging command and mains power draw command corresponding to the current time from the time period optimization results.

[0091] Specifically, the energy storage charging value, energy storage discharging value, and mains power consumption value corresponding to each time period are read from the time period optimization results in chronological order. First, the time period position corresponding to the current time is located, and the energy storage charging value and energy storage discharging value corresponding to the current time are extracted. Then, the instructions are merged according to the direction and magnitude of the energy storage charging value and energy storage discharging value to obtain the energy storage charging and discharging instruction corresponding to the current time. At the same time, the mains power consumption value corresponding to the current time is extracted and the instruction is encapsulated to obtain the mains power consumption instruction corresponding to the current time.

[0092] S6. Convert the energy storage charging and discharging commands and the mains power extraction commands into real-time control commands and execute them. After execution, collect the actual operating results and perform correction and update processing on the physical information POD modality library and the multi-timescale prediction model set based on the actual operating results, load power sequence, photovoltaic output sequence and energy storage SOC sequence.

[0093] S6.1 Configure the control content corresponding to the bidirectional DC / DC converter according to the energy storage charging and discharging command, and configure the start / stop content and power setting content corresponding to the bidirectional AC / DC converter according to the mains power supply command, and generate a real-time control command set.

[0094] Specifically, based on the charging direction, discharging direction, and power value in the energy storage charging and discharging command, the start / stop status, operating direction, and power control content of the bidirectional DC / DC converter are configured accordingly. Based on the power extraction status and power extraction power in the mains power extraction command, the start / stop content and power setting content of the bidirectional AC / DC converter are configured accordingly. The control content of the bidirectional DC / DC converter and the start / stop content and power setting content of the bidirectional AC / DC converter are merged and organized according to the execution order at the current moment to generate a real-time control command set.

[0095] S6.2 Send the real-time control command set to the corresponding execution object for control execution, and after control execution, collect the corresponding photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data and power supply matching feedback data from the multi-source operation monitoring data to obtain the actual operation results.

[0096] Specifically, real-time control commands are sent to the corresponding execution locations of the bidirectional DC / DC converter and the bidirectional AC / DC converter according to command categories. This enables the bidirectional DC / DC converter to complete the energy storage charging and discharging regulation according to the charging and discharging control content in the real-time control command set, and enables the bidirectional AC / DC converter to complete the mains power extraction regulation according to the start / stop content and power setting content in the real-time control command set. After the control execution is completed, the photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data, and power supply matching feedback data in the multi-source operation monitoring data are collected item by item at the same acquisition time as the real-time control command set. The collected photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data, and power supply matching feedback data are then uniformly merged according to time order and field order to obtain the actual operation results.

[0097] S6.3 Based on the actual operating results and the load power sequence, photovoltaic output sequence and energy storage SOC sequence, perform correlation analysis on the state changes in the power supply matching feedback data, data center load data and multi-source operation monitoring data, and generate state identification input results.

[0098] Specifically, based on the actual operating results and the load power sequence, photovoltaic output sequence, and energy storage SOC sequence, the status changes of each time period in the power supply matching feedback data, data center load data, and multi-source operation monitoring data are read and matched in a unified time sequence. The actual operating results of each time period are correlated with the load power sequence, photovoltaic output sequence, and energy storage SOC sequence of the same time period to obtain the corresponding relationship of the operating deviation of each time period. The corresponding relationship of the operating deviation of each time period is jointly merged and sorted with the power supply change information in the power supply matching feedback data, the load change information in the data center load data, and the operating status change information in the multi-source operation monitoring data, so that the power supply change, load change, and operating status change form a unified correlation result, and the status identification input result is generated.

[0099] S6.4 Based on the status recognition input results, identify the power supply deviation status and operating condition change status, and generate status recognition results.

[0100] Specifically, based on the status recognition input results, the power supply change information, load change information, and operating status change information in the status recognition input results are segmented and organized in chronological order. The power supply change information and load change information in each time period are compared to identify whether there is a deviation between the power supply capacity and the load demand, and the power supply deviation status is obtained. At the same time, the operating status change information, power supply change information, and load change information in each time period are compared in chronological order to extract the stage fluctuation characteristics and continuous change characteristics, determine the operating condition change status, and merge the power supply deviation status and operating condition change status in a unified order to generate the status recognition result.

[0101] S6.5. Based on the state identification results, perform short-term error correction processing on the power supply deviation state, and perform medium-term model update processing on the operating condition change state when the preset daily update cycle is reached, and perform long-term adaptive evolution processing on the operating condition change state when the preset long-term update cycle is reached, generating correction update processing results.

[0102] Specifically, based on the state recognition results, the deviation direction, deviation magnitude, and occurrence time corresponding to the power supply deviation state are read, and the prediction error term corresponding to the power supply deviation state is locally corrected to complete the short-term error correction processing. When the preset daily update cycle is reached, the stage change information corresponding to the operating condition change state is read, and the relevant parameters of the medium-term model in the multi-timescale prediction model set and the corresponding modal relationships in the physical information POD modal library are updated to complete the medium-term model update processing. When the preset long-term update cycle is reached, the continuous change information corresponding to the operating condition change state is read, and the relevant parameters of the long-term model in the multi-timescale prediction model set, the modal combination relationships and update rules in the physical information POD modal library are continuously adjusted to complete the long-term adaptive evolution processing. The short-term error correction processing, medium-term model update processing, and long-term adaptive evolution processing are merged in chronological order to generate the correction and update processing results.

[0103] It should also be noted that the preset daily update cycle refers to the daily update duration set in advance to ensure that the multi-timescale prediction model set and physical information POD modality library can be periodically adjusted according to recent operational changes, and is used to trigger mid-term model update processing; the preset long-term update cycle refers to the long-term update duration set in advance to reflect the continuous impact of seasonal characteristics, photovoltaic module degradation characteristics, energy storage battery aging characteristics and IT configuration replacement characteristics on the prediction process, and is used to trigger long-term adaptive evolution processing.

[0104] S6.6. Based on the correction and update processing results, collect the newly added feedback data from the actual operation results and the seasonal characteristics, photovoltaic module degradation characteristics, energy storage battery aging characteristics and IT configuration replacement characteristics from the multi-source operation monitoring data, and generate update input results.

[0105] Specifically, based on the correction and update processing results, new feedback data corresponding to short-term error correction processing, mid-term model update processing, and long-term adaptive evolution processing are read from the actual operation results. Seasonal features, photovoltaic module degradation features, energy storage battery aging features, and IT configuration replacement features corresponding to the current operation stage are extracted from multi-source operation monitoring data. The new feedback data and the seasonal features, photovoltaic module degradation features, energy storage battery aging features, and IT configuration replacement features are uniformly organized and correspondingly merged according to time order and field order. This makes the new feedback data and the seasonal features, photovoltaic module degradation features, energy storage battery aging features, and IT configuration replacement features form a set of associated data that can be used together for subsequent modal updates and model retraining, generating update input results.

[0106] S6.7. Based on the updated input results, when the preset daily update cycle is reached, the POD dimensionality reduction process of physical information constraints is re-executed to generate modal update results.

[0107] Specifically, based on the updated input results, when the preset daily update cycle is reached, the newly added feedback data, seasonal characteristics, photovoltaic module degradation characteristics, energy storage battery aging characteristics, and IT configuration replacement characteristics are first uniformly organized, and the fields are matched according to the feature types corresponding to the physical constraint condition set to obtain the updated multiphysics coupling input feature set. Based on the updated multiphysics coupling input feature set, energy conservation constraints, convection heat transfer constraints, and boundary condition constraints are reconstructed to obtain the updated physical constraint condition set. Based on the updated physical constraint condition set, singular value decomposition and dominant mode screening are re-executed on the updated multiphysics coupling input feature set, and the screened dominant modes are re-merged and organized to generate the mode update results.

[0108] S6.8. Based on the modal update results, the multi-timescale prediction model set is retrained when the preset daily update cycle is reached. When the preset long-term update cycle is reached, seasonal features, photovoltaic module degradation features, energy storage battery aging features, and IT configuration replacement features are introduced to fully retrain the multi-timescale prediction model set, resulting in the updated physical information POD modal library and multi-timescale prediction model set.

[0109] Specifically, based on the modal update results, when the preset daily update cycle is reached, the corresponding updated modalities in the modal update results are first added to the Physical Information POD modal library according to the modal sequence number. Then, the input benchmarks, network layers, and parameter update order of each prediction process in the multi-timescale prediction model set are reorganized according to the modal arrangement relationship corresponding to the updated Physical Information POD modal library. The multi-timescale prediction model set is then retrained to obtain the daily update training results. When the preset long-term update cycle is reached, the seasonal features, photovoltaic module degradation features, energy storage battery aging features, and IT configuration replacement features are merged with the daily update training results. Based on the merged long-term change relationship, each GA-BPNN prediction structure in the multi-timescale prediction model set is fully retrained. The parameter configuration after full retraining is then uniformly organized with the supplemented Physical Information POD modal library to obtain the updated Physical Information POD modal library and the multi-timescale prediction model set.

[0110] It should be noted that by forming a closed-loop collaborative link by executing real-time control commands, collecting actual operating results, and updating the physical information POD modality library and multi-timescale prediction model set, compared with the existing technology that separates prediction and control or only schedules according to fixed rules, it can continuously correct power supply deviations and improve the adaptability of multi-timescale prediction models to load fluctuations, photovoltaic changes and energy storage aging, thereby improving the accuracy, continuity and real-time response capability of prediction and control.

[0111] In summary, this invention achieves compressed representation of multi-physics coupling information by inputting the multi-physics coupled input feature set into the physical information constraint POD dimensionality reduction process. This is used to generate low-dimensional POD coefficients and a physical information POD mode library, thereby providing a stable foundation for multi-timescale prediction model sets and linear programming solutions. Furthermore, by performing correction and update processing on the physical information POD mode library and the multi-timescale prediction model set, a closed-loop connection between actual operating results and predictive control is achieved. This is used to correct power supply deviation states and operating condition changes, thereby improving the adaptability of real-time control commands and the continuity of the predictive control process.

[0112] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A predictive control method for direct optical-storage data center supply based on artificial intelligence, characterized in that: include, Data collection includes photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data, power supply matching feedback data, weather forecast data, and immersion liquid cooling operation data. Noise reduction, alignment, and normalization are performed. Based on the synergistic changes of the electric power field, thermal field, flow field, and environmental field, a multi-physics field input feature set is constructed. The POD dimensionality reduction process involves inputting a multi-physics field coupled input feature set into the physical information constraint, extracting the dominant modes under energy conservation constraints, convective heat transfer constraints, and boundary condition constraints, and obtaining low-dimensional POD coefficients through mode projection. By inputting the low-dimensional POD coefficient into a multi-timescale prediction model set, the predicted POD coefficient values ​​for multiple future time periods are obtained. Based on the predicted POD coefficient values ​​for multiple future time periods and the physical information POD mode library, the load power sequence, photovoltaic output sequence, and energy storage SOC sequence are reconstructed. A linear programming optimization model is constructed using load power sequence, photovoltaic output sequence and energy storage SOC sequence as inputs to obtain energy storage charging and discharging commands and grid power input commands. The energy storage charging and discharging commands and the mains power supply commands are converted into real-time control commands and executed. After execution, the actual operating results are collected and the physical information POD modality library and the multi-timescale prediction model set are corrected and updated based on the actual operating results, load power sequence, photovoltaic output sequence and energy storage SOC sequence.

2. The predictive control method for direct optical-storage data center supply based on artificial intelligence as described in claim 1, characterized in that: The construction of the multiphysics coupled input feature set specifically includes, Collect photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data, power supply matching feedback data, weather forecast data, and immersion liquid cooling operation data to form multi-source operation monitoring data; Multi-source operation monitoring data are uniformly aggregated and then processed by denoising, alignment and normalization to form a multi-physics field coupled input feature set; The multiphysics field includes the electric power field, thermal field, flow field, and environmental field.

3. The predictive control method for direct optical-storage data center supply based on artificial intelligence as described in claim 1, characterized in that: The obtained low-dimensional POD coefficients specifically include, Based on the feature data representing power change, heat generation change, coolant flow, heat transfer state, chip surface heat flux density and coolant inlet temperature in the multi-physics field coupled input feature set, energy conservation constraints, convective heat transfer constraints and boundary condition constraints are constructed to generate a set of physical constraint conditions. Based on the set of physical constraints, singular value decomposition is performed on the multi-physics field coupled input feature set to extract the dominant modes that satisfy the preset cumulative energy ratio and generate a dominant mode set. A physical information POD mode library is constructed based on the dominant mode set, and the multi-physics field coupled input feature set is projected onto the mode space corresponding to the physical information POD mode library to obtain low-dimensional POD coefficients.

4. The predictive control method for direct optical-storage data center supply based on artificial intelligence as described in claim 1, characterized in that: The process of obtaining the predicted POD coefficient values ​​for multiple future time periods specifically includes: According to the preset prediction task division rules, the prediction process of the low-dimensional POD coefficient is divided to generate a multi-timescale prediction process set corresponding to rapid response control, rolling coordinated adjustment and operation trend prediction respectively. Based on the multi-timescale prediction process set, and using the shared physical information POD mode basis function as a foundation, GA-BPNN prediction structures are constructed for each prediction process to generate a multi-timescale prediction model set. By inputting the low-dimensional POD coefficients into a multi-timescale prediction model set, predicted POD coefficient values ​​for multiple future time periods can be obtained.

5. The predictive control method for direct optical-storage data center supply based on artificial intelligence as described in claim 4, characterized in that: The generation of the multi-timescale prediction model set specifically includes... Extract the prediction objects and parameter update requirements corresponding to each prediction process from the multi-timescale prediction process set, and generate a task configuration subset; Based on the task configuration subset and using the shared physical information POD modal basis function as input, the network layer, node configuration, genetic algorithm optimization range and parameter update rules corresponding to each prediction process are configured to generate the prediction structure configuration result. Based on the prediction structure configuration results, a corresponding GA-BPNN prediction structure is constructed for each prediction process, and a physical driving branch is configured in each GA-BPNN prediction structure. By combining each GA-BPNN prediction structure with the physics-driven branch, a set of multi-timescale prediction models is generated.

6. The predictive control method for direct optical-storage data center supply based on artificial intelligence as described in claim 1, characterized in that: The reconstructed load power sequence, photovoltaic output sequence, and energy storage SOC sequence specifically include, Call the first K orders of physical information POD modes from the physical information POD mode library; The predicted POD coefficients for multiple future time periods are reconstructed by tensor product with the POD modes of the first K physical information to generate multi-physics collaborative reconstruction results. The corresponding reconstruction components in the multi-physics collaborative reconstruction results are classified and mapped to obtain load power sequence, photovoltaic power output sequence and energy storage SOC sequence.

7. The predictive control method for direct optical-storage data center supply based on artificial intelligence as described in claim 1, characterized in that: The requests for energy storage charging / discharging commands and mains power draw commands specifically include, Using load power sequence, photovoltaic output sequence and energy storage SOC sequence as inputs, a linear programming solution is constructed according to the optimization objectives of reducing grid electricity consumption and reducing curtailment of solar power. Based on the linear programming solution, under the constraints of power balance, energy storage SOC dynamic constraints, charging and discharging power limits, energy storage SOC safe range constraints, and mains power upper limit constraints, the energy storage charging power, energy storage discharging power, and mains power extraction power for each time period are solved to generate time period optimization results. Extract the energy storage charging / discharging command and mains power draw command corresponding to the current time from the time period optimization results.

8. The predictive control method for direct optical-storage data center supply based on artificial intelligence as described in claim 7, characterized in that: The process of converting energy storage charging / discharging commands and mains power extraction commands into real-time control commands and executing them specifically includes: Configure the corresponding control content of the bidirectional DC / DC converter according to the energy storage charging and discharging command, and configure the corresponding start / stop content and power setting content of the bidirectional AC / DC converter according to the mains power supply command to generate a real-time control command set; The real-time control command set is sent to the corresponding execution object for control execution. After the control execution, the photovoltaic array operation data, energy storage power station status data, DC bus status data, mains power standby status data, data center load data and power supply matching feedback data in the multi-source operation monitoring data are collected back to obtain the actual operation results.

9. The predictive control method for direct optical-storage data center supply based on artificial intelligence as described in claim 8, characterized in that: The process of performing correction and update on the Physical Information (POD) modality library and the multi-timescale prediction model set specifically includes: Based on the actual operating results and the load power sequence, photovoltaic output sequence and energy storage SOC sequence, the status changes in power supply matching feedback data, data center load data and multi-source operation monitoring data are correlated and analyzed to generate status identification input results. Based on the status recognition input results, the power supply deviation status and operating condition change status are identified, and status recognition results are generated. Based on the state identification results, short-term error correction processing is performed on the power supply deviation state, and medium-term model update processing is performed on the operating condition change state when the preset daily update cycle is reached. When the preset long-term update cycle is reached, long-term adaptive evolution processing is performed on the operating condition change state to generate correction update processing results.

10. The predictive control method for direct optical-storage data center supply based on artificial intelligence as described in claim 9, characterized in that: The generated correction and update processing results are used to update the Physical Information Device (POD) modality library and the multi-timescale prediction model set, specifically including: Based on the correction and update processing results, the newly added feedback data from the actual operation results and the seasonal characteristics, photovoltaic module degradation characteristics, energy storage battery aging characteristics and IT configuration replacement characteristics from the multi-source operation monitoring data are collected to generate updated input results. Based on the updated input results, the POD dimensionality reduction process with physical information constraints is re-executed when the preset daily update cycle is reached, generating modal update results; Based on the modal update results, the multi-timescale prediction model set is retrained when the preset daily update cycle is reached. When the preset long-term update cycle is reached, seasonal features, photovoltaic module degradation features, energy storage battery aging features, and IT configuration replacement features are introduced to fully retrain the multi-timescale prediction model set, resulting in the updated physical information POD modal library and multi-timescale prediction model set.

Images