Energy optimization decision system based on cloud computing

Abstract

The invention discloses an energy optimization decision-making system based on cloud computing, and belongs to the technical field of energy optimization, and the method specifically comprises the steps: collecting energy consumption data and environmental parameter data of energy consumption equipment, carrying out the preprocessing, extracting space-time joint feature vectors of the energy consumption data and the environmental parameter data of the energy consumption equipment, and carrying out the calculation of the space-time joint feature vectors; constructing a space-time collaborative prediction model of dynamic heterogeneous graph fusion, predicting energy load in a future preset time period, establishing a multi-target optimization model including economic cost, carbon emission and equipment loss, and solving the multi-target optimization model to generate an optimal scheduling instruction; according to the method, millisecond-level data analysis and scheduling optimization are realized through the dynamic heterogeneous graph fused space-time collaborative prediction model and cloud computing, and the method is suitable for large-scale complex scenes and cross-domain collaborative scheduling.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy optimization, and more specifically, it is an energy optimization decision-making system based on cloud computing. Background Art

[0002] In the field of modern energy management and optimal scheduling, how to efficiently allocate energy resources, reduce energy waste and improve energy utilization efficiency has become a research hotspot. Traditional energy scheduling methods mainly rely on centralized or distributed scheduling architectures. However, with the increasing uncertainty of energy supply and demand and the rising proportion of new energy (such as wind energy and solar energy), traditional scheduling methods have many limitations: 1. Limited computing power, making it difficult to process massive amounts of data; 2. Fixed adjustment rules, lacking the ability of dynamic optimization; 3. There are many influencing factors for the electricity load of equipment, making it difficult to predict; 4. The dilemma of large-scale optimization solution.

[0003] With the development of cloud computing, big data and artificial intelligence technologies, introducing cloud computing into the field of energy optimization scheduling, cloud computing has powerful computing power, massive data storage capacity and elastic scalability. Although it solves the problem of data processing efficiency to a certain extent, it lacks effective spatio-temporal correlation analysis, resulting in serious loss of data features and affecting the accuracy of load prediction and optimization models. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention proposes an energy optimization decision-making system based on cloud computing, which makes full use of the massive data processing ability and distributed computing advantages of cloud computing, combines spatio-temporal correlation analysis, and realizes real-time collection, intelligent analysis and scheduling of multi-source energy data.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An energy optimization decision-making system based on cloud computing, comprising: a data collection module, an energy load prediction module and an optimization decision module;

[0007] The data collection module is used to collect energy consumption data and environmental parameter data of energy-consuming equipment, and perform preprocessing;

[0008] The energy load prediction module is used to extract the spatio-temporal joint feature vectors of energy consumption data and environmental parameter data of energy-consuming equipment, construct a spatio-temporal collaborative prediction model of dynamic heterogeneous graph fusion, and predict the energy load within a preset future time period;

[0009] The optimization decision module is used to establish a multi-objective optimization model including economic cost, carbon emissions and equipment loss, and solve the multi-objective optimization model to generate an optimal scheduling instruction.

[0010] Specifically, the energy load prediction module includes: a feature extraction unit, a prediction model construction unit, and a prediction unit;

[0011] The feature extraction unit is used to extract the spatio-temporal joint feature vector of the energy consumption data of the energy-consuming equipment and the environmental parameter data;

[0012] The prediction model construction unit is used to construct a dynamic heterogeneous graph and adopt a multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion;

[0013] The prediction unit is used to input the spatio-temporal joint feature vector into the trained spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion to predict the energy load within a preset future time period.

[0014] Specifically, the extraction of the spatio-temporal joint feature vector of the energy consumption data of the energy-consuming equipment and the environmental parameter data includes:

[0015] Use bidirectional LSTM to process historical load data, capture periodic features, and output a time feature vector;

[0016] Construct a graph convolutional network, input the preprocessed energy consumption data, meteorological data, and social event data of the energy-consuming equipment at the current moment into the graph convolutional network, and extract spatial correlation through 2-layer graph convolution, that is, a spatial feature vector;

[0017] Introduce a spatio-temporal attention mechanism, calculate the interaction weight between the time feature and the spatial feature, and generate a spatio-temporal joint feature vector.

[0018] Specifically, the construction of the dynamic heterogeneous graph and the adoption of the multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion include:

[0019] Define heterogeneous nodes, dynamically model the edge relationships of the heterogeneous nodes, and construct a dynamic heterogeneous graph;

[0020] Dynamically adjust the weights of the temporal convolution;

[0021] Combine the dynamic heterogeneous graph and the multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion, and train the spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion.

[0022] Specifically, the dynamic adjustment of the weights of the temporal convolution includes:

[0023] Input the historical T-step temporal data of the node pair (v, u) in the dynamic heterogeneous graph into the depthwise separable temporal convolution network to extract cross-period features;

[0024] Set the sliding window size, splice the historical T-step time series data of the node pair (v, u), and input it into the depthwise separable time series convolutional network to generate the dynamic weight matrix W t , and update it every 5 minutes;

[0025] Define the causal constraint matrix, filter and eliminate unreasonable connections to obtain the final weight W′ t .

[0026] Specifically, the method of combining the dynamic heterogeneous graph and multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion and training the spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion includes:

[0027] Calculate the spatial correlation of node v and its neighbor node u using spatial attention, integrate the spatial correlation information of neighbor nodes, and output the spatial aggregation feature;

[0028] Calculate the importance of the time dimension of the time series feature of node v using time attention, fuse the historical time series information, and output the time aggregation feature;

[0029] Generate a gating signal through an activation function, and use the gating signal to dynamically fuse spatial and time features to obtain the final fusion feature;

[0030] Combine the dynamic heterogeneous graph and multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion, and use dynamic topology optimization for optimization training.

[0031] Specifically, the method of combining the dynamic heterogeneous graph and multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion and using dynamic topology optimization for optimization training includes:

[0032] Convert the structure of the spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion into a topological embedding vector, extract the statistical features of node degree distribution and edge weight entropy value as the input of reinforcement learning;

[0033] Define the action space, and design three types of graph structure modification operations, including: adding and deleting causal edges, adjusting the edge weight update frequency, and injecting virtual perturbation nodes;

[0034] Fuse the prediction accuracy improvement, graph structure sparsity and structural stability, construct a reward function, and drive the reinforcement learning strategy to optimize the spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion in the direction of improving prediction accuracy;

[0035] Adopt the proximal policy optimization algorithm to update the policy network offline and learn the optimal optimization strategy for the spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion.

[0036] Specifically, the optimization decision module includes: a multi-objective optimization model building unit and an optimization scheduling instruction generating unit;

[0037] The multi-objective optimization model construction unit is used to define the optimization objective function and the adaptive weight coefficient, dynamically adjust them, and construct the multi-objective optimization model;

[0038] The optimization scheduling instruction generation unit is used to input the predicted energy load data within a preset future time period into the mixed integer programming solver in the cloud, and combine the multi-objective optimization model with the real-time power grid constraints to solve the optimal scheduling solution.

[0039] Specifically, the predicted energy load data within a preset future time period is input into a mixed integer programming solver in the cloud, and the optimal scheduling solution is solved by combining a multi-objective optimization model with real-time grid constraints, including:

[0040] Deploy a parallelized mixed integer programming solver in the cloud;

[0041] The predicted energy load data for a preset future time period is input into a mixed integer programming solver in the cloud to split the global optimization problem into device-level sub-problems and system-level main problems;

[0042] Combined with the real-time power grid constraints, the mixed integer programming solver is started to solve the multi-objective optimization model in parallel and output the optimal scheduling plan.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The present invention proposes an energy optimization decision-making system based on cloud computing. Relying on the powerful computing power of cloud computing, it can process multi-energy data in real time, and adopt a parallel computing framework to achieve millisecond-level data analysis and scheduling optimization, thereby improving the timeliness and overall efficiency of energy scheduling.

[0045] 2. The present invention proposes an energy optimization decision-making system based on cloud computing. By establishing a spatiotemporal collaborative prediction model that integrates dynamic heterogeneous graphs, it can deeply mine the spatiotemporal correlation in the data and use reinforcement learning to optimize the model, greatly improving the accuracy of load forecasting.

[0046] 3. The present invention proposes an energy optimization decision-making system based on cloud computing, which is suitable for large-scale complex scenarios and cross-domain collaborative scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The cloud computing-based energy optimization decision system architecture diagram provided by the present invention;

[0048] Figure 2Schematic diagram of the process for establishing a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion provided by the present invention;

[0049] Figure 3 Flow chart of the optimization scheduling provided by the present invention. Detailed implementation manners

[0050] The following further elaborates on the present application with reference to specific embodiments. The following embodiments will assist those skilled in the art in further understanding the present application, but do not limit the present application in any way. It should be noted that those of ordinary skill in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0051] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application 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 only used to explain the present application and are not used to limit the present application.

[0052] It should be noted that if there is no conflict, the various features in the embodiments of the present application can be combined with each other, and all are within the protection scope of the present application. In addition, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flow chart, in some cases, the steps shown or described can be executed in a different order from the module division in the device or the flow chart. In addition, the terms "first", "second", "third", etc. used in the present application do not limit the data and execution order, but only distinguish the same items or similar items with basically the same functions and effects.

[0053] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used in this specification in the description of the present application are only for the purpose of describing specific embodiments and are not used to limit the present application. The term "and / or" used in this specification includes any and all combinations of one or more of the related listed items.

[0054] Embodiment 1

[0055] Please refer to Figures 1-3 , an embodiment provided by the present invention: an energy optimization decision-making system based on cloud computing, including: a data acquisition module, an energy load prediction module, and an optimization decision-making module;

[0056] The data acquisition module is used to collect energy consumption data and environmental parameter data of energy-consuming devices and perform preprocessing;

[0057] The preprocessing includes: data cleaning and spatio-temporal alignment; the data cleaning is used to remove redundant data and fill in missing data; the spatio-temporal alignment includes time granularity alignment and spatial gridification. The time granularity alignment is to uniformly convert the energy consumption data of energy-consuming devices (at a 15-minute level), meteorological data (at an hourly level, including temperature, humidity, irradiance, etc.), and social event data (unstructured text) into 5-minute time slices, and cubic spline interpolation is used to fill in the missing values of meteorological data. The spatial gridification is to divide the region into 1km×1km grids, and each grid is associated with load monitoring points, weather stations, etc., to construct a spatial adjacency matrix;

[0058] The energy load prediction module is used to extract the spatio-temporal joint feature vector of the energy consumption data of energy-consuming devices and environmental parameter data, construct a spatio-temporal collaborative prediction model based on dynamic heterogeneous graph fusion, and predict the energy load within a preset future time period;

[0059] The energy load prediction module includes: a feature extraction unit, a prediction model construction unit, and a prediction unit;

[0060] The feature extraction unit is used to extract the spatio-temporal joint feature vector of the energy consumption data of energy-consuming devices and environmental parameter data;

[0061] The specific steps for extracting the spatio-temporal joint feature vector of the energy consumption data of energy-consuming devices and environmental parameter data are as follows:

[0062] Step S1: Use bidirectional LSTM to process historical load data, capture periodic features, and output a time feature vector;

[0063] Periodic features such as weekday and weekend patterns, etc.;

[0064] Step S2: Construct a graph convolutional network, input the preprocessed energy consumption data, meteorological data, and social event data of energy-consuming devices at the current moment into the graph convolutional network, and extract spatial correlations through 2 layers of graph convolution, that is, a spatial feature vector;

[0065] A dynamic adjacency matrix is used, which is different from the traditional static distance weight, to capture the spatial impact of emergencies (such as climate anomalies) on the load;

[0066] Step S3: Introduce a spatio-temporal attention mechanism, calculate the interaction weights of time features and spatial features, and generate a spatio-temporal joint feature vector.

[0067] In this embodiment, linear transformation is performed on the time feature and the space feature to learn the internal correlation of the time feature and the influence between spatial grids, and the results of the linear transformation are added to obtain an interaction vector, which represents the coupling effect of the spatio-temporal feature. The interaction vector is normalized to obtain the interaction weights of the time feature and the space feature. The time feature and the space feature are weighted and fused to obtain a spatio-temporal joint feature vector.

[0068] The prediction model construction unit is used to construct a dynamic heterogeneous graph and adopt a multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion;

[0069] As Figure 2 shown, the steps of constructing the dynamic heterogeneous graph and adopting the multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion specifically include:

[0070] Step 1: Define heterogeneous nodes, dynamically model the edge relationships of the heterogeneous nodes, and construct a dynamic heterogeneous graph;

[0071] The heterogeneous nodes include: energy demand nodes, physical entity nodes, and virtual causal nodes;

[0072] The energy demand nodes include power load points (including historical demand time series data, geographical location coordinates, etc.); the physical entity nodes include meteorological sensors (temperature and humidity, wind speed), industrial equipment (power, operating status), and power grid topology nodes (voltage, current), etc.; the virtual causal nodes include: political and current affairs impact factors (such as electricity price adjustment, etc.), emergencies (such as typhoon warnings encoded as special nodes);

[0073] The dynamic modeling of the edge relationships of the heterogeneous nodes includes: physical connection edges, causal influence edges, and spatio-temporal association edges;

[0074] The physical connection edges include: power grid line connections, upstream and downstream supply chain relationships of equipment (fixed topology); the causal influence edges include: meteorological parameters → equipment efficiency, political and current affairs impact → industrial electricity demand (dynamic weight); the spatio-temporal association edges include: associations between regions generated based on a geographical distance attenuation function (time dependence);

[0075] The discrete signals of emergencies are encoded as temporary nodes, and dynamic causal relationship edges are established with physical entity nodes, breaking through the limitation that traditional graph structures only contain physical entities.

[0076] Step 2: Dynamically adjust the weights of the temporal convolution;

[0077] The specific steps of Step 2 are as follows:

[0078] Step 201: Input the historical T-step time series data of node pairs (v, u) in the dynamic heterogeneous graph into the depthwise separable temporal convolutional network to extract cross-cycle features;

[0079] In Step 201, the core of extracting cross-cycle features is to mine the cross-cycle dependency relationships and patterns (such as the association between meteorological-grid data in different cycles) in the historical time series data of nodes through depthwise separable temporal convolution. These features contain the association information of node pairs (v, u) in the long time series range and are the core input for subsequent calculation of dynamic weights. Without the support of cross-cycle features, dynamic weights will not be able to capture the long-term dynamic patterns of node associations;

[0080] Step 202: Set the sliding window size, splice the historical T-step time series data of node pairs (v, u), and input it into the depthwise separable temporal convolutional network to generate a dynamic weight matrix W t , and update it every 5 minutes;

[0081] In Step 202, after processing the cross-cycle features through temporal convolution, the generated dynamic weight matrix is essentially a quantization of cross-cycle associations. The dynamic weight matrix is the final result of cross-cycle feature analysis, used to characterize the dynamic changes of node relationships. It can not only integrate long-time series cycle information but also represent node associations in real time in the form of dynamic weights, ultimately enabling the graph structure to adapt to the changes in node relationships in dynamic scenarios such as typhoon passing;

[0082] Step 203: Define a causal constraint matrix, screen and eliminate unreasonable connections to obtain the final weight W′ t .

[0083] The form of the causal constraint matrix M is: M ∈ {0, 1} N×N , where N represents the total number of nodes in the dynamic heterogeneous graph, 0 represents prohibited connection, and 1 represents allowed connection, describing whether the connection between two nodes conforms to causal logic, prohibiting unreasonable connections where future events affect the past, and improving the rationality of node associations in the graph structure;

[0084] The formula for the final weight is W′2 = W t ⊙M, where ⊙ represents element-wise multiplication, that is, the elements at the corresponding positions in the matrix are multiplied. Different from ordinary matrix multiplication (row-column dot product), element-wise multiplication only screens connections and does not change the weight intensity of reasonable connections;

[0085] Compared with the static graph convolutional network, in special scenarios such as typhoon passing scenarios, the association weight between meteorological nodes and power grid nodes adaptively increases by 3.7 times, accurately reflecting the impact of special scenarios (extreme weather);

[0086] Step 3: Combine the dynamic heterogeneous graph and the multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion, and train the spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion.

[0087] The specific steps of Step 3 are as follows:

[0088] Step 301: Calculate the spatial correlation between node v and its neighbor node u using spatial attention, integrate the spatial correlation information of neighbor nodes, and output the spatial aggregation feature;

[0089] Step 302: Calculate the importance of the time dimension of the temporal feature of node v using temporal attention, fuse the historical temporal information, and output the temporal aggregation feature;

[0090] Step 303: Generate a gating signal through an activation function, and use the gating signal to dynamically fuse the spatial and temporal features to obtain the final fused feature;

[0091] Step 304: Combine the dynamic heterogeneous graph and the multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion, and perform optimization training using dynamic topology optimization.

[0092] The specific steps of Step 304 are as follows:

[0093] Step 3041: Convert the structure of the spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion into a topological embedding vector, extract statistical features such as node degree distribution and edge weight entropy value, and use them as the input of reinforcement learning;

[0094] By quantifying the graph structure features, enable the reinforcement learning model to perceive the current state of the graph and provide a basis for subsequent decisions;

[0095] Step 3042: Define the action space, and design three types of graph structure modification operations, including: adding and deleting causal edges (such as adding / deleting node associations), adjusting the edge weight update frequency (such as changing the weight update period), and injecting virtual perturbation nodes (simulating emergencies);

[0096] The three types of operations cover the connection relationship, update mechanism, and scenario simulation of the graph structure, and comprehensively support the dynamic adjustment of the graph structure;

[0097] Step 3043: Integrate prediction accuracy improvement, graph structure sparsity, and structural stability, construct a reward function, and drive the reinforcement learning strategy to optimize the spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion in the direction of improving prediction accuracy;

[0098] In step 3043, using reinforcement learning, based on the cyclic iteration of state (graph structure statistical features), action (graph structure modification operations), and reward function (prediction accuracy, computational efficiency, structural stability), the connection relationship of the graph, the weight update mechanism, and the node perturbation simulation are optimized. Finally, a graph topology structure that better fits the actual scenario (such as power grid faults) is formed. The reward function takes "prediction error reduction" as the core goal. The model processes the input data based on the optimized graph structure to achieve accurate prediction of scenarios such as before the occurrence of faults;

[0099] Step 3044: Adopt the Proximal Policy Optimization (PPO) algorithm to update the policy network offline and learn the optimization strategy of the spatio-temporal collaborative prediction model for optimal dynamic heterogeneous graph fusion.

[0100] The prediction unit is used to input the spatio-temporal joint feature vector into the trained spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion to predict the energy load within a preset future time period.

[0101] The optimization decision-making module is used to establish a multi-objective optimization model including economic cost, carbon emissions, and equipment losses, and solve the multi-objective optimization model to generate an optimal scheduling instruction.

[0102] The optimization decision-making module includes: a multi-objective optimization model construction unit and an optimal scheduling instruction generation unit;

[0103] The multi-objective optimization model construction unit is used to define the optimization objective function and the adaptive weight coefficient, dynamically adjust them, and construct a multi-objective optimization model;

[0104] The specific steps of defining the optimization objective function and the adaptive weight coefficient, and dynamically adjusting them to construct a multi-objective optimization model are as follows:

[0105] Set optimization objectives, including: minimizing energy cost, maximizing the utilization rate of renewable resources, and optimizing the stability of the power grid;

[0106] In this embodiment, minimizing the energy cost is the core economic objective. The cost items include: electricity purchase cost (time-of-use electricity price for thermal power / green power), energy storage charge and discharge loss cost (such as 5% loss for each charge and discharge of a lithium battery), equipment start-stop cost (extra fuel is required for starting and stopping a coal-fired unit), and demand-side response compensation cost (subsidy for load reduction); Scenario constraints: Charge preferentially during low electricity price periods and discharge preferentially during peak electricity price periods; Give priority to consuming photovoltaic power generation during large-scale photovoltaic generation to reduce high-price electricity purchase; Data support: Access to electricity market trading prices, equipment operation and maintenance ledgers, and historical load curve data;

[0107] Maximizing the utilization rate of renewable resources is the goal of green power consumption. The consumption rate = (actual consumption + storable amount) / total power generation. Excluding the abandoned power, such as wind turbine curtailment and PV curtailment, set the regional consumption target. For example, the provincial power grid requires the PV consumption rate ≥ 95%. Calculate the risk of abandoned power in real time. For example, when the current output exceeds the load + energy storage capacity, an alarm is triggered. Data support: PV / wind power prediction data, remaining capacity of energy storage, real-time value of grid load;

[0108] Optimizing the grid stability is the safety goal. Key indicators: bus voltage deviation (±5% is normal), line load rate (warning when exceeding 80%), frequency fluctuation (±0.2Hz is the limit). Convert the grid power flow calculation results into dispatching restrictions. For example, when a certain line is overloaded, it is prohibited to increase the transmission power of this line; Data support: voltage, current, and frequency data collected by the grid SCADA system in real time, and the health status of equipment (such as the aging degree of the transformer affects the load upper limit).

[0109] Build a dynamic weight adjustment trigger logic, including primary trigger, secondary trigger, and tertiary trigger, and establish a scenario-index-weight mapping table, using fuzzy logic to handle boundary conditions;

[0110] In this embodiment, the primary trigger (conventional scenario): preset the basic weight according to the time period. For example, during weekdays, cost is emphasized during the day, and renewable energy consumption is emphasized at night. Exemplarily, during the morning peak (7 - 9 o'clock), the electricity price is high, and the cost weight is 60%; during the late night (0 - 5 o'clock), PV does not generate electricity, and the wind power consumption weight is 70%. The secondary trigger (event-driven): dynamically adjust the weight according to real-time events (such as when the grid load rate > 85%, the stability weight rises to 50%). Exemplarily, a typhoon causes multiple line trips, triggering the "stability first" mode, automatically freezing the energy storage charge and discharge plan, and giving priority to ensuring the grid frequency. The tertiary trigger (extreme scenario): set a hard constraint threshold (such as when the abandoned power rate > 10%, force the renewable energy weight to 100%, allowing the cost to rise briefly). Exemplarily, when PV generates a large amount of power and the load is at a low valley, and the energy storage is full, start the emergency abandoned power plan, and at the same time trigger the demand response to invite users to increase the load temporarily;

[0111] Use fuzzy logic to handle boundary conditions. Exemplarily, when "load rate 75% - 85%" and "renewable energy output exceeds the prediction by 20%", calculate through fuzzy rules: stability weight = base value (30%) + 0.5 × the over-limit ratio of (85% - 75%) = 35%, renewable energy weight = base value (40%) + 0.3 × the over-output ratio = 46%, and the cost weight automatically drops to 19%.

[0112] Based on the optimization goal and dynamic weight, build a multi-objective optimization model.

[0113] In this embodiment, the multi-objective optimization model is the sum of "optimization objectives × dynamic weights of each objective".

[0114] The optimization scheduling instruction generation unit is configured to input the predicted energy load data within a preset future time period into a mixed-integer programming solver in the cloud, and combine the multi-objective optimization model with real-time power grid constraints to solve the optimal scheduling plan.

[0115] As Figure 3 shown, the steps of inputting the predicted energy load data within a preset future time period into a mixed-integer programming solver in the cloud, combining the multi-objective optimization model with real-time power grid constraints, and solving the optimal scheduling plan specifically include:

[0116] Step A1: Deploy a parallelized mixed-integer programming solver in the cloud;

[0117] In this embodiment, first, the hardware resources are pooled, and 50 high-performance servers (each with 32-core CPUs + 256 GB of memory) are deployed. Through Kubernetes container management, 3 resource groups are divided: a fast-solving group (10 units): to handle emergency scheduling (such as re-optimization within 15 minutes after a fault), and the computing power is preferentially allocated; a regular-solving group (30 units): to handle the rolling scheduling 4 times a day (0:00 / 6:00 / 12:00 / 18:00); a standby group (10 units): to cope with sudden load surges or solver failures and take over tasks in real time; configure a high-speed shared storage (NVMe cluster) to store power grid topology data, device parameters, and historical scheduling plans, with a latency < 1 ms.

[0118] Using mixed-integer programming for decomposition, splitting in the time dimension: The 24-hour scheduling cycle is sliced by 15 minutes to generate 96 sub-problems of time periods (T1 - T96), and each sub-problem includes:

[0119] Decision variables: start-stop status of thermal power units, charge-discharge power of energy storage, cross-regional power transmission plan;

[0120] Constraint transfer: The SOC of the energy storage and the operating status of the unit in the previous time period are used as the initial conditions for the next time period.

[0121] Splitting in the space dimension: According to the power grid partitions (such as three regions A / B / C), each region independently solves the sub-problem, and synchronizes the power limit of the tie line through "boundary constraints";

[0122] For example: When the photovoltaic power generation is large in area A, the sub-problem needs to synchronously update the power transmission upper limit to area B to avoid line overload.

[0123] Parallel strategy, adopting the "master-slave architecture": 1 master server is responsible for task distribution and result aggregation; 49 slave servers solve sub-problems in parallel; pre-load common constraint templates (such as the minimum start-stop time of equipment and the charge-discharge rate limit of energy storage) to reduce repeated calculations.

[0124] Step A2: Input the predicted energy load data within a preset future time period into the mixed-integer programming solver in the cloud, and split the global optimization problem into device-level sub-problems and system-level master problems.

[0125] Step A3: Combine the real-time grid constraint conditions, start the mixed-integer programming solver, perform parallel solution on the multi-objective optimization model, and output the optimal scheduling plan.

[0126] In this embodiment, the parallelization efficiency is optimized and fault tolerance is improved by dynamically monitoring the computing power of the slave servers: when the CPU utilization rate of a certain server > 90%, automatically allocate subsequent sub-problems to low-load nodes; give priority to processing sub-problems during "critical periods" (such as peak load periods T28 - T32) to ensure the accuracy of peak scheduling; fault tolerance and retry mechanism: when the solution of a sub-problem times out (> 15 minutes), automatically switch the solver (such as switching from CPLEX to Gurobi) and reuse the calculated feasible solutions; if it still fails, mark this period as "manual intervention" and generate a temporary instruction based on the historical similar-day plan; data exception handling: if the topological data of a certain regional power grid is missing, automatically load the backup data of the previous 1 hour, attach a "data missing warning", and through parallel solution, achieve the efficient solution of mixed-integer programming in power grid scheduling.

[0127] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are the same as the corresponding technical solutions in the prior art in terms of implementation principles are not described in detail to avoid excessive elaboration.

[0128] As described above in the specific implementation manner, the purpose, technical solution and beneficial effects of the present invention are further described in detail. It should be understood that the above is only the specific implementation manner of the present invention and is not used to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An energy optimization decision-making system based on cloud computing, characterized in that, Including: A data acquisition module, an energy load prediction module, and an optimization decision-making module; The data acquisition module is used to collect energy consumption data and environmental parameter data of energy-consuming devices and perform preprocessing; The energy load prediction module is used to extract spatio-temporal joint feature vectors of energy consumption data and environmental parameter data of energy-consuming devices, construct a spatio-temporal collaborative prediction model based on dynamic heterogeneous graph fusion, and predict the energy load within a preset future time period; The optimization decision-making module is used to establish a multi-objective optimization model including economic cost, carbon emissions, and equipment loss, and solve the multi-objective optimization model to generate an optimal scheduling instruction.

2. The energy optimization decision-making system based on cloud computing according to claim 1, characterized in that, The energy load prediction module includes: a feature extraction unit, a prediction model construction unit, and a prediction unit; The feature extraction unit is used to extract spatio-temporal joint feature vectors of energy consumption data and environmental parameter data of energy-consuming devices; The prediction model construction unit is used to construct a dynamic heterogeneous graph and adopt a multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model based on dynamic heterogeneous graph fusion; The prediction unit is used to input the spatio-temporal joint feature vector into the trained spatio-temporal collaborative prediction model based on dynamic heterogeneous graph fusion to predict the energy load within a preset future time period.

3. The energy optimization decision-making system based on cloud computing according to claim 2, characterized in that, The extraction of the spatio-temporal joint feature vector of the energy consumption data and environmental parameter data of the energy-consuming devices includes: Using bidirectional LSTM to process historical load data, capture periodic features, and output a time feature vector; Constructing a graph convolutional network, inputting the preprocessed energy consumption data, meteorological data, and social event data of the energy-consuming devices at the current moment into the graph convolutional network, and extracting spatial correlation through 2-layer graph convolution, that is, a spatial feature vector; Introducing a spatio-temporal attention mechanism to calculate the interaction weights between the time feature and the spatial feature, and generating a spatio-temporal joint feature vector.

4. The energy optimization decision-making system based on cloud computing according to claim 2, wherein The construction of the dynamic heterogeneous graph and the adoption of the multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model based on dynamic heterogeneous graph fusion include: Defining heterogeneous nodes and dynamically modeling the edge relationships of the heterogeneous nodes to construct a dynamic heterogeneous graph; Dynamically adjusting the weights of the temporal convolution; Combining the dynamic heterogeneous graph and the multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model based on dynamic heterogeneous graph fusion, and training the spatio-temporal collaborative prediction model based on dynamic heterogeneous graph fusion.

5. The energy optimization decision-making system based on cloud computing according to claim 4, characterized in that, The dynamic adjustment of the weights of the temporal convolution includes: Inputting the historical T-step temporal data of the node pair (v, u) in the dynamic heterogeneous graph into a depthwise separable temporal convolutional network to extract cross-period features; Set the sliding window size, splice the historical T-step time series data of the node pair (v, u), and input it into the depthwise separable temporal convolutional network to generate the dynamic weight matrix W t , and update it every 5 minutes; Define the causal constraint matrix, screen and eliminate unreasonable connections to obtain the final weight W t '.

6. The energy optimization decision-making system based on cloud computing according to claim 4, wherein The combination of the dynamic heterogeneous graph and the multi-modal attention fusion mechanism to establish a spatio-temporal collaborative prediction model based on dynamic heterogeneous graph fusion, and the training of the spatio-temporal collaborative prediction model based on dynamic heterogeneous graph fusion include: Using spatial attention to calculate the spatial correlation between node v and its neighbor node u, integrating the spatial correlation information of the neighbor nodes, and outputting a spatial aggregation feature; Using time attention to calculate the importance of the time dimension of the temporal feature of node v, and fusing historical temporal information to output a time aggregation feature; Generating a gating signal through an activation function, and dynamically fusing the spatial and temporal features using the gating signal to obtain a final fusion feature; Combining a dynamic heterogeneous graph and a multi-modal attention fusion mechanism, a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion is established and optimized and trained using dynamic topology optimization.

7. The energy optimization decision-making system based on cloud computing according to claim 6, wherein The above-mentioned combining a dynamic heterogeneous graph and a multi-modal attention fusion mechanism, establishing a spatio-temporal collaborative prediction model for dynamic heterogeneous graph fusion, and optimizing and training using dynamic topology optimization includes: Transform the spatio-temporal collaborative prediction model structure of dynamic heterogeneous graph fusion into a topological embedding vector, and extract node degree distribution and edge weight entropy value statistical features as the input of reinforcement learning. Define the action space and design three types of graph structure modification operations, including: adding or deleting causal edges, adjusting the edge weight update frequency, and injecting virtual perturbation nodes. Fusing prediction accuracy improvement, graph structure sparsity, and structural stability, construct a reward function to drive the reinforcement learning strategy to optimize the spatio-temporal collaborative prediction model of dynamic heterogeneous graph fusion in the direction of improving prediction accuracy. Adopt the proximal policy optimization algorithm to offline update the policy network and learn the optimal optimization strategy for the spatio-temporal collaborative prediction model of dynamic heterogeneous graph fusion.

8. The energy optimization decision-making system based on cloud computing according to claim 7, characterized in that, The optimization decision module includes: a multi-objective optimization model construction unit and an optimization scheduling instruction generation unit. The multi-objective optimization model construction unit is used to define the optimization objective function and the adaptive weight coefficient, dynamically adjust them, and construct a multi-objective optimization model. The optimization scheduling instruction generation unit is used to input the predicted energy load data within a preset future time period into the mixed integer programming solver in the cloud, and combine the multi-objective optimization model with the real-time power grid constraint conditions to solve the optimal scheduling plan.

9. The energy optimization decision-making system based on cloud computing according to claim 8, wherein, The above-mentioned inputting the predicted energy load data within a preset future time period into the mixed integer programming solver in the cloud, combining the multi-objective optimization model with the real-time power grid constraint conditions, and solving the optimal scheduling plan includes: Deploy a parallelized mixed integer programming solver in the cloud. Input the predicted energy load data within a preset future time period into the mixed integer programming solver in the cloud, and split the global optimization problem into device-level sub-problems and system-level main problems. Combined with the real-time power grid constraint conditions, start the mixed integer programming solver to perform parallel solution on the multi-objective optimization model and output the optimal scheduling plan.

Images