Energy equipment intelligent regulation and control method and system based on Internet of Things

Abstract

The invention discloses an energy equipment intelligent regulation and control method and system based on the Internet of Things, and the method comprises the steps: collecting multi-source heterogeneous data containing equipment operation state data and environment parameters, and generating a standardized equipment operation feature vector; inputting the equipment operation feature vector into a federated learning-based regional energy load prediction model, and generating a partition load prediction curve in a future preset time period; according to the partition load prediction curve, combined with equipment operation constraint conditions, adopting a multi-objective optimization algorithm to dynamically solve optimal operation parameters of each energy equipment, and generating an equipment regulation and control instruction set; and based on the device regulation and control instruction set, automatically verifying and executing the regulation and control instruction through the smart contract execution layer enabled by the block chain. By utilizing the embodiment of the invention, the unification of the regional energy equipment in three dimensions of data privacy security, accurate and efficient regulation and control and credible instruction execution can be realized, and safe, economic and self-consistent intelligent regulation and control are achieved.

Description

Technical Field

[0001] This invention belongs to the field of energy equipment technology, and in particular to an intelligent control method and system for energy equipment based on the Internet of Things. Background Technology

[0002] With the deepening application of IoT technology in the energy sector, the regulation and control of regional energy systems faces new challenges. Currently, traditional centralized regulation methods rely on uploading all operational data from each node to a cloud center for processing. This approach not only demands high network bandwidth and suffers from data transmission delays, but more importantly, the aggregation of large amounts of raw data involves user privacy and trade secrets, posing data security risks. Furthermore, existing methods often focus on a single objective (such as economic efficiency) when formulating regulation strategies, making it difficult to achieve dynamic and coordinated optimization of multiple objectives such as energy efficiency, stability, and load balancing while meeting equipment operational constraints. In addition, the execution of regulation commands lacks transparency and immutability, posing a risk of malicious attacks or misoperation. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent control method and system for energy equipment based on the Internet of Things (IoT) to address the shortcomings of existing technologies. This system can achieve a unified approach to regional energy equipment in three dimensions: data privacy and security, precise and efficient control, and reliable command execution, thereby achieving safe, economical, and self-consistent intelligent control.

[0004] One embodiment of this application provides a smart control method for energy devices based on the Internet of Things, the method comprising: Multi-source heterogeneous data containing equipment operating status data and environmental parameters is collected by a multi-modal sensor array deployed on energy equipment nodes, and edge computing nodes are used to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data to generate standardized equipment operating feature vectors. The device operation feature vector is input into a regional energy load prediction model based on federated learning. By fusing the features of multi-node data without exchanging the original data, a regional load prediction curve for a future preset period is generated. Based on the zonal load forecast curves and combined with equipment operating constraints, a multi-objective optimization algorithm is used to dynamically solve for the optimal operating parameters of each energy device, and generate a set of equipment control instructions. Based on the set of equipment control commands, the control commands are automatically verified and executed through a blockchain-enabled smart contract execution layer, thereby achieving distributed collaborative optimization and safe management of regional energy equipment.

[0005] Optionally, the step of collecting multi-source heterogeneous data including equipment operating status data and environmental parameters through a multi-modal sensor array deployed on energy equipment nodes, and using edge computing nodes to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data to generate a standardized equipment operating feature vector, includes: By synchronously collecting equipment operating status data and environmental parameters through a multimodal sensor array deployed at energy equipment nodes, a raw multi-source heterogeneous dataset is generated. The original multi-source heterogeneous dataset is time-stamp aligned and spatially calibrated. An interpolation algorithm is used to compensate for the sampling time difference of different sensors, generating a spatiotemporally aligned multi-source data sequence. Based on spatiotemporally aligned multi-source data sequences, key operational features are extracted using feature extraction algorithms to generate multi-dimensional feature vectors. The multidimensional feature vectors are standardized, and the Z-score normalization method is used to eliminate the difference in dimensions, finally generating standardized equipment operation feature vectors.

[0006] Optionally, the step of inputting the device operating feature vector into a regional energy load forecasting model based on federated learning, and generating a zoned load forecast curve for a future preset time period by fusing multi-node data features without exchanging the original data, includes: Initialize federated learning clients at each edge computing node, load pre-trained load prediction models, and input standardized device running feature vectors into the local model for forward computation to generate local model update gradients. Homomorphic encryption is used to encrypt the gradients of the local model update. The encrypted gradients are then transmitted to the federated learning server via a secure communication protocol to generate an encrypted gradient set. On the federated learning server, secure aggregation calculations are performed on the encrypted gradient set, and the federated averaging algorithm is used to fuse the model updates of each node to generate global model update parameters. The regional energy load forecasting model is updated based on the global model update parameters, and the updated model is used to predict the load for a preset period in the future, ultimately generating the zonal load forecasting curve.

[0007] Optionally, the step of dynamically solving for the optimal operating parameters of each energy device using a multi-objective optimization algorithm based on the zonal load forecast curve and combined with equipment operating constraints, and generating a set of equipment control instructions, includes: Analyze the load distribution characteristics in the zoned load forecast curves, extract peak load periods, load change trends, and load spatial distribution information, and generate a load characteristic analysis report; Based on the load characteristic analysis report, a multi-objective optimization model considering equipment operation constraints is established. The constraints include upper and lower limits of equipment power, operating efficiency range, and equipment start-up and shutdown limits, generating a mathematical model of the optimization problem. The mathematical model of the optimization problem is solved by using a multi-objective particle swarm optimization algorithm, which simultaneously optimizes multiple objectives including energy utilization efficiency, equipment operating cost and system stability, and generates the optimal solution set. The Pareto optimal solution that meets the actual engineering requirements is selected from the set of optimal solutions, analyzed into specific equipment operating parameters, and finally a set of equipment control instructions is generated.

[0008] Optionally, the step of automatically verifying and executing control commands based on the set of device control commands through a blockchain-enabled smart contract execution layer to achieve distributed collaborative optimization and security management of regional energy equipment includes: Encode the set of equipment control instructions into a data structure that can be recognized by smart contracts to generate standardized control instruction data packets; The standardized control instruction data packets are broadcast to each energy equipment node through the blockchain network, triggering the automatic verification logic of the smart contract and generating instruction verification results. Based on the instruction verification results, the smart contract automatically executes the verified control instructions, issues specific control commands through the device control interface, and generates device control signals; The system monitors the equipment's execution status in real time and records the results to a blockchain distributed ledger, ensuring the traceability and immutability of the control process, and ultimately achieving distributed collaborative optimization and safe management of regional energy equipment.

[0009] Another embodiment of this application provides an intelligent control system for energy equipment based on the Internet of Things, the system comprising: The acquisition module is used to acquire multi-source heterogeneous data containing equipment operating status data and environmental parameters through a multi-modal sensor array deployed on energy equipment nodes, and to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data using edge computing nodes to generate standardized equipment operating feature vectors. The prediction module is used to input the device operation feature vector into the regional energy load prediction model based on federated learning, and generate the zonal load prediction curve for a future preset period by fusing the data features of multiple nodes without exchanging the original data. The generation module is used to dynamically solve the optimal operating parameters of each energy device based on the partition load prediction curve and combined with the equipment operation constraints, and generate a set of equipment control instructions. The control module is used to automatically verify and execute the control commands based on the set of equipment control commands through a blockchain-enabled smart contract execution layer, thereby realizing distributed collaborative optimization and safety management of regional energy equipment.

[0010] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0011] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0012] Compared with existing technologies, this invention provides an intelligent control method for energy equipment based on the Internet of Things (IoT). It collects multi-source heterogeneous data, including equipment operating status data and environmental parameters, through a multimodal sensor array deployed on energy equipment nodes, generating standardized equipment operating feature vectors. These feature vectors are then input into a regional energy load prediction model based on federated learning to generate regional load prediction curves for a preset future time period. Based on the regional load prediction curves and combined with equipment operating constraints, a multi-objective optimization algorithm dynamically solves for the optimal operating parameters of each energy device, generating a set of equipment control instructions. Based on this set of instructions, a blockchain-enabled smart contract execution layer automatically verifies and executes the control instructions, thereby achieving a unified approach to regional energy equipment control in three dimensions: data privacy and security, precise and efficient control, and reliable instruction execution. This results in safe, economical, and self-consistent intelligent control. Attached Figure Description

[0013] Figure 1 A hardware structure block diagram of a computer terminal for an intelligent control method for energy equipment based on the Internet of Things provided in an embodiment of the present invention; Figure 2 A flowchart illustrating an intelligent control method for energy equipment based on the Internet of Things (IoT) provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent control system for energy equipment based on the Internet of Things, provided as an embodiment of the present invention. Detailed Implementation

[0014] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] This invention first provides an intelligent control method for energy equipment based on the Internet of Things. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0016] The following detailed explanation uses a computer terminal as an example. Figure 1This is a hardware structure block diagram of a computer terminal for an intelligent control method for energy equipment based on the Internet of Things, provided as an embodiment of the present invention. (See diagram below.) Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0017] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any IoT-based intelligent control method for energy devices.

[0018] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0019] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any kind of intelligent control method for energy devices based on the Internet of Things.

[0020] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0021] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0022] See Figure 2 The present invention provides an intelligent control method for energy equipment based on the Internet of Things, which may include the following steps: S201: Multi-source heterogeneous data containing equipment operating status data and environmental parameters are collected by a multi-modal sensor array deployed on energy equipment nodes, and the multi-source heterogeneous data is spatiotemporally aligned and feature extracted by edge computing nodes to generate a standardized equipment operating feature vector. Specifically, multimodal sensor arrays deployed on energy equipment nodes can be used to synchronously collect equipment operating status data and environmental parameters to generate raw multi-source heterogeneous datasets; This step is the foundational data acquisition stage for intelligent control of energy equipment. Its core is to comprehensively capture key information about equipment operation and the environment through multiple types of sensors, ensuring that the data covers the core dimensions required for equipment status monitoring and load forecasting, while also guaranteeing the synchronization of data acquisition and the integrity of the raw data. The specific implementation method is as follows: The deployment of multimodal sensor arrays needs to be customized for different types of energy equipment, covering typical energy equipment nodes such as photovoltaic panels, energy storage batteries, charging piles, and distributed wind turbines. Each equipment node should be equipped with at least 6 types of sensors to form a full-dimensional acquisition system of "operating status + environmental parameters". Operational status data acquisition focuses on core equipment performance indicators: For photovoltaic panels, voltage sensors (measurement range 0-1000V, accuracy ±0.5%), current sensors (0-50A, accuracy ±0.3%), and power sensors (0-50kW, accuracy ±0.2%) are deployed to capture output voltage, output current, and instantaneous power in real time; for energy storage batteries, voltage sensors (0-500V), current sensors (-50A-50A, supporting bidirectional charging and discharging measurement), temperature sensors (-20℃-85℃, accuracy ±0.2℃), and SOC (State of Charge) sensors (0%-100%, accuracy ±1%) are deployed to record individual cell voltage, charging and discharging current, battery temperature, and remaining capacity; for charging piles, voltage sensors (0-750V), current sensors (0-300A), and power factor sensors (0.8-1.0, accuracy ±0.01) are deployed to collect output voltage, charging current, and power factor.

[0023] Environmental parameter acquisition targets key external factors affecting equipment operating efficiency, with each equipment node equipped with a light intensity sensor (0-2000W / m²). 2 The system includes sensors for various parameters, such as: an ambient temperature sensor (-40℃-85℃, accuracy ±0.3℃), a wind speed sensor (0-25m / s, accuracy ±0.2m / s), and a humidity sensor (0%-100%RH, accuracy ±3%RH). The photovoltaic panel area focuses on enhancing the collection of light intensity and ambient temperature data, while the wind turbine area emphasizes wind speed and direction data (with an additional wind direction sensor deployed, 0°-360°, accuracy ±5°). The charging pile area is supplemented with a rainfall sensor (0-50mm / h, accuracy ±0.1mm / h) to address the impact of severe weather on equipment operation.

[0024] Synchronous data acquisition from sensors is achieved through clock synchronization commands from edge computing nodes. All sensors are connected to a local industrial Ethernet network, and the clock is calibrated using the NTP (Network Time Protocol) v4 protocol, with synchronization accuracy controlled within ±1ms, ensuring that all sensors on the same device node acquire data at the same timestamp. The sampling frequency is dynamically adjusted according to device characteristics: the sampling frequency for rapidly changing parameters such as voltage and current is set to 10Hz (acquiring data once every 100ms), while the sampling frequency for slowly changing parameters such as temperature, SOC, and light intensity is set to 1Hz (acquiring data once every 1 second). Derivative parameters such as power and power factor are calculated and generated in real time by the edge computing nodes (based on synchronous voltage and current data).

[0025] The original multi-source heterogeneous dataset is stored in a structured format of "Device ID-Timestamp-Parameter Type-Value-Unit-Sensor ID". For example, a data record for a photovoltaic panel node is "PV-001, 2025-06-10 14:30:00.000, Output Voltage, 380.5V, SENSOR-V-01", and a data record for an energy storage battery is "BAT-003, 2025-06-10 14:30:00.000, SOC, 78.2%, SENSOR-SOC-03". The dataset supports real-time writing to the local storage module of the edge computing node (storage capacity ≥1TB, read / write speed ≥500MB / s), and data quality labels ("valid", "invalid", "suspicious") are added. When a sensor value exceeds the normal range (e.g., a voltage sensor collects 1200V) or the value fluctuates by more than 10% for three consecutive collections, it is marked as "suspicious", providing a quality basis for subsequent data processing.

[0026] The original multi-source heterogeneous dataset is time-stamp aligned and spatially calibrated. An interpolation algorithm is used to compensate for the sampling time difference of different sensors, generating a spatiotemporally aligned multi-source data sequence. This step is crucial for addressing the heterogeneity of multi-sensor data. It eliminates spatiotemporal biases by using time synchronization and spatial correlation, and then compensates for sampling frequency differences through interpolation, ensuring that all data can be fused and analyzed within a unified spatiotemporal framework. The specific implementation method is as follows: The timestamp alignment is based on the unified clock of the edge computing node. Based on the synchronization result of the NTPv4 protocol, the timestamps of the original data are corrected. For the time differences caused by inconsistent sampling frequencies (such as a 10Hz voltage sensor and a 1Hz temperature sensor), the "reference time axis interpolation" strategy is adopted: taking the timestamps of the highest sampling frequency (10Hz) as the reference, a continuous time axis (time interval 100ms) is constructed. For the data of sensors with low sampling frequencies, between the adjacent actual sampling timestamps, data is supplemented at the reference time points through interpolation algorithms to ensure that there is complete device status and environmental parameter data at each reference time point.

[0027] The linear interpolation algorithm is selected for interpolation, which is applicable to parameters with gentle numerical changes (such as temperature, SOC, light intensity, etc.). The calculation formula is y = y1+(x - x1)×(y2 - y1) / (x2 - x1), where x is the target reference timestamp, x1 and x2 are the two nearest actual sampling timestamps before and after this target timestamp (x1 < x < x2), y1 and y2 are the sensor values corresponding to the timestamps, and y is the interpolation result of the target timestamp. For example, the temperature sensor collects 28.5°C at 14:30:00.000 and 28.7°C at 14:30:01.000. Then the interpolated temperature at the reference timestamp 14:30:00.100 is 28.5+(0.1 - 0.0)×(28.7 - 28.5) / (1.0 - 0.0)=28.52°C, and the interpolated temperature at 14:30:00.200 is 28.54°C, and so on, to complete the temperature data of 10 reference time points. For parameters such as current and voltage that fluctuate rapidly but without sudden changes, the cubic spline interpolation algorithm is adopted. By constructing a smooth interpolation curve, the continuity and rationality of the supplemented data are ensured, and the errors caused by linear interpolation are avoided.

[0028] The spatial position calibration aims to establish the association between sensor data and the physical positions of devices. The "global coordinate + local coordinate" dual calibration strategy is adopted: all device nodes obtain global coordinates (accuracy ±2m) through GPS modules. For example, the coordinates of the PV-001 node in the photovoltaic panel array are (116.397°, 39.908°), and the coordinates of the BAT-003 node in the energy storage battery pack are (116.398°, 39.909°); at the same time, a local coordinate system is established within the regional energy system (with the regional control center as the origin, the x-axis in the east-west direction, the y-axis in the north-south direction, and the unit in meters), and the global coordinates are converted into local coordinates. For example, the local coordinates of PV-001 are (50.2m, 30.5m), ensuring that the data can be clustered and analyzed according to spatial positions. Each sensor data is bound to the corresponding device node coordinates and the description of the sensor installation position (such as "left side of the upper surface of PV-001", "cell No. 1 of the cell group of BAT-003"), realizing the full-link association of "data - sensor - device - position".

[0029] The multi-source data sequence generated after spatiotemporal alignment is indexed by a reference time axis. Each timestamp corresponds to a complete data record containing 20+ dimensions of parameters. For example, the sequence record for 14:30:00.000 is: "Timestamp: 2025-06-10 14:30:00.000; Device ID: PV-001; Local coordinates: (50.2, 30.5); Output voltage: 380.5V; Output current: 12.3A; Instantaneous power: 4.68kW; Battery temperature: 28.5℃; Illumination intensity: 1200W / m²". 2 Ambient temperature: 26.3℃; Humidity: 45%RH… The temporal continuity and spatial correlation of the data sequence meet the requirements for subsequent feature extraction.

[0030] Based on spatiotemporally aligned multi-source data sequences, key operational features are extracted using feature extraction algorithms to generate multi-dimensional feature vectors. This step is the core of transforming raw data into features that the model can recognize. By extracting time-domain, frequency-domain, and device-specific features, key information in the data is condensed, redundancy is eliminated, and a multi-dimensional feature vector that reflects the correlation between device operating status and load is generated. The specific implementation method is as follows: The feature extraction algorithm employs a multi-dimensional extraction strategy combining time-domain statistical features, frequency-domain features, and device-specific features to ensure comprehensiveness and relevance. Time-domain statistical features focus on the numerical distribution and trends of the data, calculating statistics for each parameter at 100 consecutive reference time points (corresponding to 10 seconds of data): mean (reflecting the average level of the parameter), variance (reflecting the degree of fluctuation), peak value (maximum value), trough value (minimum value), peak-to-peak value (difference between peak and trough values), skewness (reflecting the symmetry of the distribution), and kurtosis (reflecting the steepness of the distribution). For example, in 10 seconds of photovoltaic panel output current data (100 sampling points), the mean is 12.5A, and the variance is 0.3A. 2 The peak current is 13.2A, the valley current is 11.8A, the peak-to-peak current is 1.4A, the skewness is 0.2 (approximately symmetrical distribution), and the kurtosis is 2.8 (close to normal distribution). These statistical characteristics can effectively reflect the stability of the current.

[0031] Frequency domain features target the periodic fluctuations during equipment operation. A Fast Fourier Transform (FFT) is used to convert time-domain data into frequency-domain data, extracting features such as the fundamental frequency, harmonic content, and total harmonic distortion (THD). For example, after FFT processing, the time-domain data of the charging pile's output voltage shows a fundamental frequency of 50Hz (power frequency), a 3rd harmonic content of 1.2%, a 5th harmonic content of 0.8%, and a total harmonic distortion of 1.5%. These features reflect the degree of voltage waveform distortion and indirectly relate to equipment operating efficiency. The calculation window for frequency domain features is set to 1 second (10 sampling points), with a sliding step size of 0.5 seconds to ensure the capture of frequency changes within short periods.

[0032] The equipment's unique features are customized based on the operating mechanisms of different energy devices, highlighting core indicators strongly correlated with load forecasting and control optimization: Photovoltaic panel photoelectric conversion efficiency (instantaneous power / (illuminance × cell area)) is extracted, with the cell area preset to 2m². 2 Maximum power point tracking (MPPT) efficiency (actual output power / theoretical maximum output power); SOC change rate (change in SOC per unit time, e.g., 0.5% / min) and charge / discharge efficiency (discharge power / charging power × 100%) for energy storage batteries, and remaining cycle life percentage (calculated based on the number of charge / discharge cycles and design cycle life); charging rate (charged amount per unit time, e.g., 2kW·h / min) and power factor deviation (actual power factor - target power factor 0.95) for charging piles; wind energy utilization coefficient (actual captured power / theoretical maximum wind power) and tip speed ratio (tip linear velocity / wind speed) for distributed wind turbines.

[0033] The construction of multidimensional feature vectors integrates all extracted features in the order of "device type + feature category". The dimension of each feature vector is adjusted according to the device type. The feature vector dimension of photovoltaic panel equipment is 32 dimensions, energy storage battery is 36 dimensions, charging pile is 30 dimensions, and distributed wind turbine is 34 dimensions. Taking photovoltaic panels as an example, the feature vector is composed as follows: time-domain statistical features (16 dimensions including mean voltage, voltage variance, peak voltage, mean current, current variance, peak current, mean power, power variance, peak power, mean temperature, temperature variance, mean illuminance, peak illuminance, etc.), frequency-domain features (4 dimensions including voltage THD, current THD, fundamental frequency deviation, etc.), and device-specific features (12 dimensions including photoelectric conversion efficiency, MPPT efficiency, power fluctuation coefficient, etc.). The final generated multi-dimensional feature vector example is "[380.2,0.8,385.5,12.4,0.3,13.2,4.72,0.15,4.98,28.3,0.5,1180,1350,...,0.85,0.92,0.02]". Each dimension of the feature corresponds to a clear physical meaning, providing high-quality data support for subsequent model input.

[0034] The multidimensional feature vectors are standardized, and the Z-score normalization method is used to eliminate the difference in dimensions, finally generating standardized equipment operation feature vectors.

[0035] This step is crucial for eliminating interference from different feature units. By using Z-score normalization, all features are mapped to a uniform numerical range, ensuring balanced weights for each feature during model training and inference, and avoiding model bias caused by differences in units. The specific implementation method is as follows: The core principle of Z-score normalization is to convert the original value of each feature into a standard normal distribution Z-score. The calculation formula is z=(x-μ) / σ, where x is the original value of the feature, μ is the mean of the feature in the historical dataset (feature data in the most recent 72 hours, sample size ≥10000), σ is the standard deviation of the feature, and z is the normalized feature value. The mean of the normalized data is 0, the standard deviation is 1, and the value range is usually concentrated between [-3,3]. Outliers outside this range will be marked and treated as ±3 (to avoid the influence of extreme values).

[0036] The calculation of the mean μ and standard deviation σ requires a sufficient sample size of historical data to ensure the reliability of the statistical results. For example, in the historical data of the photovoltaic panel's "photovoltaic conversion efficiency" feature, there are 12,000 samples, and the calculated mean μ = 0.82 and standard deviation σ = 0.05; the historical data of the "mean output voltage" feature has a mean μ = 380V and a standard deviation σ = 5V; the historical data of the "mean light intensity" feature has a mean μ = 1000W / m 2 Standard deviation σ = 200 W / m 2 These statistical parameters need to be updated in real time and recalculated every 24 hours to adapt to changes in equipment operating status and environmental conditions.

[0037] Standardization requires performing Z-score calculations on each feature value in the multidimensional feature vector dimension by dimension. For example, in the multidimensional feature vector of a photovoltaic panel, the original value of "photovoltaic conversion efficiency" is 0.85, and after normalization, z = (0.85 - 0.82) / 0.05 = 0.6; the original value of "average output voltage" is 385V, and after normalization, z = (385 - 380) / 5 = 1.0; the original value of "average light intensity" is 1300W / m². 2 After normalization, z = (1300-1000) / 200 = 1.5; the original value of "voltage THD" is 1.2%, the historical mean μ = 1.0%, the standard deviation σ = 0.3%, and after normalization, z = (1.2-1.0) / 0.3 ≈ 0.67. For outliers exceeding the range [-3,3], such as a sudden change in light intensity to 2000 W / m² at a certain moment... 2After normalization, z = (2000-1000) / 200 = 5.0, which exceeds the upper limit of 3.0. Therefore, it is processed as 3.0 to ensure the stability of the data.

[0038] The resulting standardized equipment operation feature vector has values ​​for all dimensions distributed within the range of [-3,3] with no unit difference. For example, the standardized feature vector for a photovoltaic panel is "[1.0,0.4,1.1,0.8,0.6,1.2,0.96,0.3,1.04,0.4,0.2,0.9,1.75,...,0.6,0.8,0.04]". This vector retains the relative differences and physical meaning of the original features while eliminating dimensional interference, allowing it to be directly input into the regional energy load prediction model, ensuring that the model can fairly utilize the features across all dimensions for learning and prediction.

[0039] S202, the device operation feature vector is input into the regional energy load prediction model based on federated learning, and the regional load prediction curve for the future preset period is generated by fusing the data features of multiple nodes without exchanging the original data. Specifically, federated learning clients can be initialized at each edge computing node, pre-trained load prediction models can be loaded, and standardized device running feature vectors can be input into the local model for forward computation to generate local model update gradients. This step is the starting point for federated learning's distributed training. Its core is to utilize local data to complete local model updates without leaking the original data of each node. This is achieved by initializing the client, loading the adapted model, and performing forward computation and gradient calculation to generate local model update gradients, providing a foundation for global model optimization. The specific implementation is as follows: The initialization of the federated learning client for edge computing nodes needs to match the node deployment of the regional energy system. Assuming there are 10 edge computing nodes (covering different energy equipment clusters such as photovoltaic arrays, energy storage clusters, and charging pile clusters), each node deploys an independent client instance. The client has built-in communication, model calculation, and data security modules. During initialization, the client establishes an encrypted connection with the federated learning server through a secure bootstrap protocol to complete identity authentication (using a public-key-based digital signature mechanism; each client has a unique device certificate, and the server establishes a connection after verifying the certificate's validity), ensuring the trustworthiness of both communicating parties.

[0040] The pre-trained load prediction model uses a Long Short-Term Memory (LSTM) network to adapt to the time-series prediction characteristics of energy load. The model structure includes an input layer (32-dimensional, consistent with the standardized equipment operation feature vector), two hidden layers (containing 128 and 64 neurons respectively), a dropout layer (dropout rate of 0.2 to prevent overfitting), and an output layer (1-dimensional, outputting the load prediction value for the next hour). The model pre-training data comes from historical publicly available load data of the regional energy system (excluding privacy data of each node). After pre-training, the prediction error (MAE) on the validation set is 3.2%, meeting the basic requirements for initial training. When each client loads the model, the initial weight parameters of the model are obtained synchronously (e.g., the weight matrix from the input layer to the first hidden layer has a dimension of 32×128, and the bias vector has a dimension of 128), ensuring that the initial model parameters of all clients are completely consistent.

[0041] After the standardized equipment operation feature vector is input into the local model, forward computation is performed: the feature vector (32-dimensional, such as [1.0, 0.4, 1.1, ..., 0.04]) is passed into the first hidden layer through the input layer, and a 128-dimensional hidden layer output is obtained by calculating through the ReLU activation function; this output is passed into the second hidden layer, and processed by the LSTM gating mechanism (forget gate, input gate, output gate) to capture temporal dependencies and output a 64-dimensional feature; after being randomly deactivated by 20% of neurons in the dropout layer, it is passed into the output layer and a linear activation function is used to obtain the predicted value (such as 45.6kW, representing the predicted load of the area corresponding to the node in the next hour).

[0042] The local model update gradient is generated based on backpropagation of the loss function. The loss function used is mean squared error (MSE), calculated as L = (y_pred - y_true)^2, where y_pred is the model's predicted load, and y_true is the historical actual load of the region corresponding to that node (local privacy data, not transmitted externally). The partial derivatives of the loss function with respect to each weight parameter of the model are calculated using the backpropagation algorithm (chain rule), i.e., the gradient. For example, the gradient from the input layer to the weight matrix of the first hidden layer is a 32×128 matrix, where each element represents the contribution of that weight to the loss; the gradient of the bias vector is a 128-dimensional vector. After the gradient calculation is completed, the client performs preliminary preprocessing on the gradient (removing NaN values, limiting the gradient magnitude to the range of [-10, 10] to avoid gradient explosion), and generates local model update gradients. For example, the gradient set of a certain client is "weight gradient W1: 32×128 matrix, bias gradient b1: 128-dimensional vector, weight gradient W2: 128×64 matrix, bias gradient b2: 64-dimensional vector, output layer weight gradient W3: 64×1 matrix, bias gradient b3: 1-dimensional vector".

[0043] Homomorphic encryption is used to encrypt the gradients of the local model update. The encrypted gradients are then transmitted to the federated learning server via a secure communication protocol to generate an encrypted gradient set. This step is crucial for ensuring data privacy. Homomorphic encryption is used to encrypt local gradients, ensuring that no node's privacy information is leaked during transmission and aggregation. Simultaneously, a secure communication protocol ensures the security of the transmission link. The specific implementation is as follows: The homomorphic encryption technology uses a partially homomorphic encryption algorithm, supporting gradient addition (meeting the aggregation requirements of the federated averaging algorithm). The encryption process includes three core steps: key generation, encryption, and decryption. First, the federated learning server generates a public-key and private-key pair. The public key is distributed to all edge computing clients, while the private key is exclusively kept by the server (used for subsequent decryption of the aggregated gradient). The key length is set to 2048 bits to ensure encryption strength (difficulty to crack ≥ 10^120 operations). After obtaining the public key, the client performs an encryption operation on each element of the gradient update of the local model. The encryption formula is E(x) = (x × g^r) mod p, where x is the gradient element value, g is the encryption primitive (preset to 2), r is a random number (1 ≤ r ≤ p - 2), and p is a large prime number (2048 bits). This formula maps the gradient element to an encrypted large integer. For example, if a gradient element value is 0.8, the encrypted result is E(0.8) = 23456789... (a 2048-bit large integer). The original gradient information is hidden and can only be decrypted using the private key.

[0044] Encryption processing must balance security and computational efficiency. Gradient matrices and vectors are encrypted element-wise. For example, a 32×128 weight gradient matrix requires encrypting 4096 elements. Encryption time for each element is ≤1ms, and the overall encryption time for a single client is ≤50ms, meeting real-time requirements. After encryption, the client organizes all encrypted gradient elements into an encrypted gradient data packet according to the structure "model parameter identifier - encrypted value," for example, "W1_0_0:23456789...,W1_0_1:12345678...,...,b3:98765432...". The data packet includes the client ID, timestamp, and data verification code (generated using the SHA-256 hash algorithm to ensure data transmission integrity).

[0045] The secure communication protocol uses TLS 1.3 to establish an end-to-end encrypted transmission channel between the client and the federated learning server. During transmission, the AES-256-GCM encryption algorithm is used to encrypt data packets twice, and a perfect forward secrecy (PFS) mechanism is enabled. Each session generates an independent session key, ensuring the security of historical session data even if the private key is leaked long-term. A timeout retransmission mechanism (500ms timeout) and flow control (single packet size ≤ 1MB to avoid network congestion) are implemented during transmission. Each client's encrypted gradient data packet is transmitted to the server through this channel. Upon receiving the packet, the server verifies the data checksum. If the verification passes, the data is stored in the encrypted gradient buffer; if the verification fails, the client is requested to retransmit.

[0046] Once all 10 edge computing nodes have transmitted and verified their encrypted gradient data packets, the federated learning server integrates them to generate an encrypted gradient set. The set is stored in categories according to client ID, such as "client 1 encrypted gradient packet, client 2 encrypted gradient packet, ..., client 10 encrypted gradient packet". Each packet contains complete encrypted gradient information for the corresponding node, preparing for subsequent secure aggregation.

[0047] On the federated learning server, secure aggregation calculations are performed on the encrypted gradient set, and the federated averaging algorithm is used to fuse the model updates of each node to generate global model update parameters. This step is the core aggregation step in federated learning. It eliminates the risk of privacy leakage from single-node gradients through secure aggregation, and uses a federated averaging algorithm to fuse gradients from multiple nodes to generate global model update parameters, ensuring the globality and accuracy of model optimization. The specific implementation is as follows: The core of secure aggregation computation is to perform gradient addition within the encrypted domain, without decrypting the gradient of individual nodes, thus avoiding privacy leaks. The federated learning server first groups the encrypted gradient elements corresponding to each model parameter in the encrypted gradient set. For example, it groups the encrypted gradient elements of all clients' "W1_0_0" into one group, the encrypted gradient elements of "W1_0_1" into another group, and so on, forming a total of (32×128+128+128×64+64+64×1+1)=4096+128+8192+64+64+1=12545 groups.

[0048] Homomorphic addition is performed on each group of encrypted gradient elements. Utilizing the homomorphic nature of addition in homomorphic encryption, the sum of all encrypted elements in the group is calculated: E(x1) + E(x2) + ... + E(xn) = E(x1 + x2 + ... + xn), where n is the number of clients (10), and x1-xn are the original values ​​of the gradient elements at each node. The addition operation is performed within the encrypted domain, and the server can never obtain the specific value of a single xi. For example, if a group of encrypted gradient elements is E(0.8), E(0.6), ..., E(0.7), the addition operation yields E(0.8 + 0.6 + ... + 0.7) = E(7.2), thus achieving encrypted aggregation of multi-node gradients.

[0049] After aggregation, the server uses a private key kept exclusively in its custody to decrypt the encrypted aggregation result, obtaining the sum of each group of gradient elements. For example, decrypting E(7.2) yields 7.2, which is the sum of the original values ​​of the gradient elements at the 10 nodes. The decryption process requires strict access control; only the server's core encryption module can access the private key. The decrypted gradient sum is stored in encrypted memory to avoid the risk of leakage caused by disk storage.

[0050] The FedAvg algorithm is used to fuse the sum of decrypted gradients to generate global model update parameters. Its core principle is to assign weights to each node based on its data volume; nodes with larger data volumes have higher gradient weights, ensuring the model update more closely matches the data distribution. The weight calculation formula is wi = ni / N, where wi is the weight of the i-th node, ni is the local training sample size of the i-th node (e.g., client 1 has 1000 samples, client 2 has 1200 samples, ..., client 10 has 900 samples), and N is the total sample size of all nodes (1000 + 1200 + ... + 900 = 10500). For example, the weight w1 of client 1 is approximately 0.095 (1000 / 10500), and the weight w2 of client 2 is approximately 0.114 (1200 / 10500).

[0051] The global model update parameters are calculated as a weighted sum of the gradients of each node and their corresponding weights, i.e., the global gradient G = w1 × g1 + w2 × g2 + ... + wn × gn, where g1 - gn is the complete set of gradients after decryption for each node. For example, the weight gradient W1 from the input layer to the first hidden layer has a global gradient W1_global = w1 × W1_1 + w2 × W1_2 + ... + w10 × W1_10, where W1_1 is the W1 gradient matrix for client 1, W1_2 is the W1 gradient matrix for client 2, and so on. Through this calculation, the model update information of all nodes is fused to generate global model update parameters, including global weight gradients W1_global, W2_global, and W3_global, and global bias gradients b1_global, b2_global, and b3_global, ensuring that the model update reflects the data features of all nodes.

[0052] The regional energy load forecasting model is updated based on the global model update parameters, and the updated model is used to predict the load for a preset period in the future, ultimately generating the zonal load forecasting curve.

[0053] This step is the final and application stage of federated learning training. It optimizes model weights by updating parameters globally, and then uses the optimized model to perform multi-time period and regional load forecasting, generating intuitive forecast curves to provide a basis for subsequent equipment control decisions. The specific implementation method is as follows: The regional energy load forecasting model is updated using the stochastic gradient descent (SGD) optimization algorithm with a learning rate of 0.001 (balancing convergence speed and stability). The update formula is θ_new = θ_old - η × G, where θ_new represents the updated model parameters, θ_old represents the original model parameters, η is the learning rate, and G represents the global model update parameters. For example, the weight matrix θ_old (32 × 128) from the input layer to the first hidden layer is subtracted by 0.001 × W1_global (global weight gradient) to obtain the updated weight matrix θ_new; the bias vector b1_old (128-dimensional) is subtracted by 0.001 × b1_global (global bias gradient) to obtain the updated bias vector b1_new. All model parameters (W1, b1, W2, b2, W3, b3) are updated according to this formula to complete model optimization.

[0054] The updated model needs to undergo performance validation. The server obtains a small amount of validation data (such as de-identified load trend data) from each client without privacy information. The prediction error is calculated on the validation set. If the MAE is ≤ 3.0%, the model update is considered valid; if the MAE is > 3.0%, the learning rate is adjusted (reduced to 0.0005) and the model is updated again until the performance requirements are met. For example, if the validation MAE of the model after a certain round of updates is 2.8%, the update is valid and can be used for subsequent predictions.

[0055] The projected time period is set to 24 hours, divided into 24 prediction periods with a 1-hour time granularity (00:00-01:00, 01:00-02:00, ..., 23:00-24:00). The zoning rules are based on the grid topology and equipment distribution of the regional energy system, dividing the area covered by the 10 edge computing nodes into 3 load zones (Zone 1: Photovoltaic + Charging Pile Cluster, Zone 2: Energy Storage + Residential Electricity Cluster, Zone 3: Industrial Load + Distributed Wind Turbine Cluster). Prediction for each zone is collaboratively completed by the corresponding client cluster.

[0056] During the forecasting process, each client inputs its latest standardized equipment operating feature vector (data updated in real-time within the last 24 hours) into the updated model, performing forward calculations to obtain the load forecast values ​​for each time period of the corresponding partition. For example, clients 1-4 in partition 1 output their forecast values ​​for each time period. The server averages the forecast values ​​from all clients in the same partition to obtain the partition load forecast value (reducing single-node forecasting errors). For example, for the time period of 08:00-09:00 in partition 1, client 1 forecasts 52.3kW, client 2 forecasts 53.1kW, client 3 forecasts 51.8kW, and client 4 forecasts 52.6kW. After averaging, the predicted load for partition 1 during this time period is 52.4kW.

[0057] The load forecast curves for each zone are generated with time (0-24 hours) on the horizontal axis and load (kW) on the vertical axis. Each zone corresponds to one curve, and the curves are processed using smooth interpolation (cubic spline interpolation) to ensure trend continuity. For example, the forecast curve for zone 1 shows a low load (30-40kW) from 0-6 hours, a rapid increase (40-52kW) from 6-8 hours, a high load (52-55kW, peak load 54.8kW, corresponding to the 10:00-11:00 period) from 8-12 hours, a slow decrease (55-45kW) from 12-18 hours, and a continuous decrease (45-32kW) from 18-24 hours. The peak load of the curve for zone 2 occurs between 19:00-20:00 (68.5kW), and the peak load of the curve for zone 3 occurs between 14:00-15:00 (85.2kW). The curve simultaneously marks the peak load period, peak load value, and average load value (average 42.3kW for zone 1, 51.7kW for zone 2, and 68.9kW for zone 3), generating a complete zone load prediction curve, providing accurate load basis for subsequent multi-objective optimization.

[0058] S203, Based on the zonal load prediction curve and combined with the equipment operation constraints, a multi-objective optimization algorithm is used to dynamically solve the optimal operating parameters of each energy device and generate a set of equipment control instructions; Specifically, it can analyze the load distribution characteristics in the zoned load forecast curve, extract peak load periods, load change trends and load spatial distribution information, and generate a load characteristic analysis report; This step is a prerequisite for multi-objective optimization. Its core is to extract key load characteristics strongly correlated with equipment regulation from the prediction curves, transforming abstract curve data into quantifiable feature indicators. This provides a clear basis for establishing subsequent optimization models. The specific implementation method is as follows: The load forecast curves for each zone contain load data for the next 24 hours for three load zones (Zone 1: PV + charging pile cluster, Zone 2: Energy storage + residential electricity cluster, Zone 3: Industrial load + distributed wind turbine cluster). The analysis process needs to be carried out according to three dimensions: "time period characteristics - trend characteristics - spatial characteristics". Peak load period extraction focuses on the maximum load value and corresponding time period of each zone. This is determined by traversing the curve data: the peak load of Zone 1 is 54.8kW, corresponding to the time period of 10:00-11:00 (morning peak charging pile usage); the peak load of Zone 2 is 68.5kW, corresponding to the time period of 19:00-20:00 (evening peak residential electricity consumption); and the peak load of Zone 3 is 85.2kW, corresponding to the time period of 14:00-15:00 (peak industrial production load). Simultaneously, the valley load and time period are extracted: valley load 28.3kW (03:00-04:00) for zone 1, valley load 35.7kW (02:00-03:00) for zone 2, and valley load 52.1kW (01:00-02:00) for zone 3. The peak-valley difference (zone 1: 26.5kW, zone 2: 32.8kW, zone 3: 33.1kW) is calculated to reflect the intensity of load fluctuation.

[0059] Load change trends are quantified by calculating the load change rate, which is calculated using the formula r = (P2 - P1) / (t2 - t1), where P1 and P2 are the load values ​​for adjacent time periods, and t1 and t2 are the start times of the time periods (in hours). For example, in zone 106:00-08:00, the load increases from 32.1kW to 52.4kW, with a change rate r = (52.4 - 32.1) / (8 - 6) = 10.15kW / h (rapid increase); from 12:00-18:00, the load decreases from 54.2kW ​​to 45.3kW, with a change rate r = (45.3 - 54.2) / (18 - 12) = -1.48kW / h (slow decrease). Trend levels are classified according to the absolute value of the rate of change: |r|≥8kW / h is "rapid change", 3kW / h≤|r|<8kW / h is "moderate change", and |r|<3kW / h is "stable change". Based on this, the trend type of each time period is determined. For example, in zone 2, the rate of change is 12.3kW / h (rapid increase) from 18:00 to 19:00, and the rate of change is -4.2kW / h (moderate decrease) from 19:00 to 21:00.

[0060] Load spatial distribution information is analyzed through load proportion and coordination relationship of each zone: the total load (P_total=P1+P2+P3) and the proportion of each zone (α1=P1 / P_total, α2=P2 / P_total, α3=P3 / P_total) are calculated for any time period. For example, from 10:00 to 11:00, the total load is 208.5kW, α1=54.8 / 208.5≈26.3%, α2=68.5 / 208.5≈32.9%, α3=85.2 / 208.5≈40.8%, with the industrial load having the highest proportion; from 03:00 to 04:00, the total load is 116.1kW, α1=28.3 / 116.1≈24.4%, α2=35.7 / 116.1≈30.7%, α3=52.1 / 116.1≈44.9%, with the industrial load still having the highest proportion, but the overall load decreases. The analysis of synergistic relationships focuses on the complementarity of loads between zones. For example, the peak load of zone 1 (photovoltaic) from 10:00 to 11:00 overlaps with the peak industrial load of zone 3, which needs to be balanced through energy storage. The peak load of zone 2 from 19:00 to 20:00 complements the valley period of zone 1, and the pressure can be alleviated by charging piles during off-peak hours.

[0061] The final load characteristic analysis report includes a summary of core indicators, trend charts, and control recommendations: the core indicators clearly define the peak and valley periods, load values, peak-valley differences, and key period change rates and proportions for each zone; the trend charts indicate the 24-hour trend level distribution for each zone; the control recommendations point out that "from 10:00 to 11:00, it is necessary to focus on increasing the energy storage discharge power to balance the peak load of zones 1 and 3, and from 19:00 to 20:00, it is necessary to guide the charging piles in zone 1 to stagger their peak hours until after 23:00," providing targeted directions for optimizing the model.

[0062] Based on the load characteristic analysis report, a multi-objective optimization model considering equipment operation constraints is established. The constraints include upper and lower limits of equipment power, operating efficiency range, and equipment start-up and shutdown limits, generating a mathematical model of the optimization problem. This step is the core framework construction stage of multi-objective optimization. By defining the optimization objective, decision variables, and constraints, the equipment control problem is transformed into a solvable mathematical model, ensuring that the optimization result meets both load requirements and equipment operating rules. The specific implementation method is as follows: First, the optimization objectives are clearly defined, focusing on the core needs of the regional energy system. Three mutually constraining objective functions are set, all presented in a maximization or minimization form for easy algorithm solution. The first objective is to maximize energy utilization efficiency (η), where η = regional renewable energy consumption / total load consumption × 100%. Renewable energy consumption includes the power generation of photovoltaic panels and distributed wind turbines, and total load consumption is the sum of the loads of the three zones. The goal is to improve the utilization rate of clean energy and reduce fossil fuel supplementation. The second objective is to minimize equipment operating costs (C), where C = Σ (equipment unit energy consumption cost × operating energy consumption + start-up and shutdown cost × number of start-ups and shutdowns). Photovoltaic panels and wind turbines have no fuel cost, and the unit energy consumption cost is 0.05 yuan / kWh (maintenance cost). The unit charging cost of energy storage batteries is 0.3 yuan / kWh, and the discharging cost is 0.1 yuan / kWh. The unit operating cost of charging piles is 0.2 yuan / kWh. The equipment start-up and shutdown cost is calculated at 5 yuan / time for energy storage batteries and 1 yuan / time for charging piles. The goal is to control operation and maintenance expenses. The third objective is to maximize system stability (S), where S = 1 - (standard deviation of load fluctuation in each period / rated total load). The rated total load is set at 300kW. The smaller the standard deviation of load fluctuation, the higher the system stability. The goal is to avoid equipment shocks caused by sudden increases or decreases in load.

[0063] The decision variables are defined as the key operating parameters of each energy device, covering all controllable objects: photovoltaic panel output power (P_pv, unit kW), distributed wind turbine output power (P_wind, unit kW), energy storage battery charging and discharging power (P_bat, charging is positive, discharging is negative, unit kW), and total charging power of charging piles (P_ev, unit kW). The variables need to be optimized according to time periods, with each time period corresponding to a set of variable values, for a total of 24 sets of decision variables over 24 hours.

[0064] The constraints are set based on the physical characteristics of the equipment and the actual engineering situation to ensure the feasibility of the optimization results. Equipment power upper and lower limits constraints: Photovoltaic panel P_pv∈[0,50]kW (rated power 50kW), wind turbine P_wind∈[0,30]kW (rated power 30kW), energy storage battery P_bat∈[-40,40]kW (rated charge / discharge power 40kW), charging pile P_ev∈[0,60]kW (rated total power 60kW). Operating efficiency range constraints: Photovoltaic panel η_pv≥80% (actual efficiency = output power / theoretical maximum power, theoretical maximum power is positively correlated with light intensity), wind turbine η_wind≥75%, energy storage battery charge / discharge comprehensive efficiency η_bat≥70%, charging pile η_ev≥90%. Equipment start / stop restrictions: The interval between two consecutive start / stop operations of the energy storage battery must be ≥30 minutes (to avoid frequent start / stop damage to the battery); the charging time for a single charging pile must be ≥15 minutes; the number of charging piles started / stopped at the same time must be ≤30% of the total number of charging piles (20 total charging piles, i.e., ≤6). In addition, a power balance constraint is added: P_pv + P_wind + P_bat - P_ev = P_total (total load for each time period) to ensure supply and demand balance.

[0065] The final generated mathematical model of the optimization problem is presented in the complete form of "objective function - decision variables - constraints". For example, the objective function set is maxη, minC, maxS, the decision variables are P_pv(t), P_wind(t), P_bat(t), P_ev(t) (t=1-24 hours), and the constraints include the mathematical expressions of the above four types of constraints, providing a clear computational framework for subsequent algorithm solutions.

[0066] The mathematical model of the optimization problem is solved by using a multi-objective particle swarm optimization algorithm, which simultaneously optimizes multiple objectives including energy utilization efficiency, equipment operating cost and system stability, and generates the optimal solution set. This step is the core of solving the optimization model. The Multi-Objective Particle Swarm Optimization (MOPSO) algorithm finds the Pareto optimal solution in the solution space of multi-objective conflict by simulating the cooperation and evolution of the particle swarm, taking into account the optimization requirements of the three objectives. The specific implementation is as follows: The core principle of the multi-objective particle swarm optimization algorithm is to treat each potential solution as a "particle". The particle updates its position and velocity in the solution space by tracking the individual optimal solution and the global optimal solution, and gradually approaches the optimal region. First, the particles are encoded. Each particle corresponds to a set of decision variables for 24 hours. The encoding length is 4×24=96 dimensions (4 device parameters × 24 time periods). For example, a part of the encoding of a particle is "P_pv(1)=32kW,P_wind(1)=15kW,P_bat(1)=8kW,P_ev(1)=25kW,P_pv(2)=35kW,...". The value of each dimension of the particle must meet the constraint condition (such as P_pv(1)∈[0,50]kW). During initialization, 100 particles are generated by random sampling to form the initial population.

[0067] The fitness function is used to evaluate the quality of particles. Since the three objectives conflict with each other (such as increasing efficiency may increase costs), a weighted summation method is used to transform the multi-objective fitness value into a single objective. The weights are set based on the priority of the regional energy system: energy utilization efficiency weight w1=0.4 (policy-oriented clean energy consumption), operating cost weight w2=0.3 (economic demands), and system stability weight w3=0.3 (safety demands). The fitness function is F=w1×η_norm+w2×(1-C_norm)+w3×S_norm, where η_norm, C_norm, and S_norm are the normalized results of the objective values ​​(mapped to the [0,1] interval). The larger the F value, the better the particle.

[0068] The particle update rule includes velocity update and position update. The velocity update formula is v_i(t+1)=ω×v_i(t)+c1×r1×(pbest_i-x_i(t))+c2×r2×(gbest-x_i(t)), where ω is the inertia weight (controlling the influence of historical velocity, with an initial value of 0.9, which linearly decreases to 0.4 during iteration to balance global exploration and local development), c1 and c2 are learning factors (both set to 2.0, controlling the intensity of particle learning towards individual and global optima, respectively), r1 and r2 are random numbers in the range [0,1], pbest_i is the individual optimal position of the i-th particle (the encoding corresponding to its own historical optimal fitness), gbest is the global optimal position (the encoding corresponding to the population's historical optimal fitness), and x_i(t) is the current position of the i-th particle. The position update formula is x_i(t+1)=x_i(t)+v_i(t+1). After the update, the values ​​of each dimension of the particle need to be constrained and verified. If the values ​​exceed the upper and lower limits of the parameters (such as P_pv=55kW), the values ​​are truncated to the boundary value (50kW) to ensure that the particle is always in the feasible solution space.

[0069] The iterative process sets a maximum of 200 iterations. Every 10 iterations, non-dominated solutions (i.e., Pareto optimal solutions, where no other solution is better than all objectives) are recorded in the population. Crowding density is used to maintain solution diversity (avoiding solutions concentrating in local regions). For example, after 50 iterations, non-dominated solutions A (η=88%, C=0.75 yuan / kWh, S=0.92), B (η=85%, C=0.70 yuan / kWh, S=0.90), and C (η=90%, C=0.80 yuan / kWh, S=0.88) appear in the population. These solutions constitute the optimal solution set, with each solution corresponding to a different set of objective trade-offs, providing diverse options for subsequent selections.

[0070] The Pareto optimal solution that meets the actual engineering requirements is selected from the set of optimal solutions, analyzed into specific equipment operating parameters, and finally a set of equipment control instructions is generated.

[0071] This step is crucial for implementing the optimization results. By setting engineering screening criteria, practical solutions are selected from Pareto optimal solutions. The abstract particle code is transformed into specific parameters that the equipment can execute, generating standardized control commands. The specific implementation method is as follows: The solutions in the Pareto optimal solution set have advantages for different objectives. Screening criteria should be set according to actual engineering needs to ensure that the selected solution is both feasible and practical: energy utilization efficiency η≥85% (meeting the requirements for clean energy consumption), operating cost C≤0.8 yuan / kWh (controlling economic expenditure), system stability S≥0.85 (ensuring safe operation of equipment), and the energy storage battery should be required to have ≤8 start-stop times per day (extending service life), and the charging power of the charging pile during peak hours (10:00-11:00, 19:00-20:00) should be ≤40kW (avoiding grid overload).

[0072] The optimal solution set is traversed according to the screening criteria. For example, solution A (η=88%, C=0.75 yuan / kWh, S=0.92) satisfies all criteria and shows a balance between efficiency and stability, so it is selected as the final practical solution. The particle code of this solution is parsed into specific equipment operating parameters, expanded according to "equipment type-time period-parameter value-constraint satisfaction": the photovoltaic panel outputs 48kW (≤50kW, η=89%≥80%) from 10:00-11:00, and the output power linearly increases from 32kW to 45kW from 06:00-08:00 (matching the load increase trend); the energy storage battery discharges 30kW (≤40kW, η_bat=72%≥7) from 10:00-11:00. 0%), charging power is 25kW from 03:00 to 04:00, with 6 start-stop cycles per day (≤8 times); charging piles have a charging power of 38kW (≤40kW, η_ev=92%≥90%) from 19:00 to 20:00, and a charging power of 55kW (off-peak charging) from 23:00 to 00:00; distributed wind turbines have an output power of 28kW (≤30kW, η_wind=78%≥75%) from 14:00 to 15:00, and the output power is dynamically adjusted according to wind speed at other times.

[0073] The equipment control instruction set is generated in a structured format of "equipment ID-control period-operating parameters-execution requirements". Each equipment corresponds to multiple instructions. For example, the instruction for photovoltaic panel ID: PV-001 is "control period: 10:00-11:00; operating parameters: output power 48kW, MPPT efficiency ≥89%; execution requirements: real-time monitoring of light intensity, power fluctuation ≤±2kW"; the instruction for energy storage battery ID: BAT-001 is "control period: 10:00-11:00; operating parameters: discharge power 30kW, charge and discharge efficiency ≥72%; execution requirements: battery temperature controlled at 25-35℃, single discharge duration ≥15 minutes"; the instruction for charging pile cluster ID: EV-001 is "control period: 19:00-20:00; operating parameters: total charging power 38kW, power distribution evenly among individual piles; execution requirements: number of piles started and stopped ≤6, charging voltage stable at 380±5V".

[0074] The instruction set includes a checksum and execution priority (level 1 for peak load periods, level 2 for stable periods, and level 3 for off-peak periods) to ensure that equipment executes according to priority. It also specifies the tolerance range for parameter adjustments (e.g., power parameters are allowed to fluctuate by ±5%), providing flexibility for equipment execution. The final generated equipment control instruction set covers the 24-hour operating plan for all energy equipment and can be directly interfaced with the subsequent smart contract execution layer, providing precise execution basis for distributed collaborative optimization.

[0075] S204, based on the set of equipment control instructions, the control instructions are automatically verified and executed through the blockchain-enabled smart contract execution layer, thereby realizing distributed collaborative optimization and safe management of regional energy equipment.

[0076] Specifically, the set of equipment control instructions can be encoded into a data structure that can be recognized by smart contracts to generate standardized control instruction data packets; This step is crucial for achieving compatibility between instructions and smart contracts. The core is to transform disparate device control instructions into a structured, parsable data format using unified encoding rules. This ensures that smart contracts can accurately identify the instruction content, the executor, and the constraints. The specific implementation method is as follows: The equipment control instruction set includes 24-hour operating parameters for various energy devices such as photovoltaic panels, energy storage batteries, charging piles, and distributed wind turbines. Each instruction contains five core elements: "equipment identifier - control period - operating parameters - execution constraints - priority". Before coding, the compatible data format of the smart contract must be clearly defined, and a lightweight JSON structure should be selected (balancing parsing efficiency and data integrity). Field definitions should follow the principles of "simplicity + standardization" to avoid ambiguity.

[0077] The coding rules must cover all instruction elements, specifically defined as follows: "device_id" is the unique identifier for the device (consisting of an 8-digit alphanumeric combination; the first two digits represent the device type, such as PV for photovoltaic panels and BAT for energy storage batteries; the last six digits are the serial number, such as PV000001 and BAT000003); "control_period" is the control period (format: "YYYY-MM-DDHH:MM-HH:MM", accurate to the minute, such as "2025-06-11 10:00-11:00"); "operation_params" is the dictionary of operating parameters, with the key being the parameter name (e.g., "out"). "put_power" represents output power, and "charge_discharge_power" represents charging and discharging power. The value is the parameter value plus the unit (the value is rounded to one decimal place, and the unit is abbreviated in English, such as "48.0kW" or "30.0kW"); "execution_constraints" is a list of execution constraints, including upper and lower limits of parameters, efficiency requirements, etc. (such as "output_power≥0.0kW&&≤50.0kW" or "efficiency≥80%"); "priority" is the execution priority (levels 1-3, with level 1 being the highest, corresponding to peak load period instructions).

[0078] Taking the instruction for photovoltaic panel PV000001 as an example, the encoded JSON structure is: {“device_id”:“PV000001”,“control_period”:“2025-06-1110:00-11:00”,“operation_params”:{“output_power”:“48.0kW”,“mppt_efficiency”:“89.0%”},“execution_constraints”:[“output_power≥0.0kW&&≤50.0kW”,“mppt_efficiency≥80%”],“priority”:1}. The instruction code for energy storage battery BAT000003 is: {“device_id”:“BAT000003”,“control_period”:“2025-06-11 10:00-11:00”,“operation_params”:{“charge_discharge_power”:“-30.0kW” (the negative sign represents discharge),“battery_temperature”:“28.5℃”},“execution_constraints”:[“charge_discharge_power≥-40.0kW&&≤40.0kW”,“battery_temperature≥20.0℃&&≤35.0℃”,“cycle_count≤8”],“priority”:1}.

[0079] After encoding, the JSON data needs to be compressed and verified: the GZIP compression algorithm is used to compress the data volume to 60% of the original size (reducing blockchain storage and transmission overhead). A data verification code (32-byte string, such as "7a9f3d8c...2b7e") is generated using the SHA-256 hash algorithm. The verification code is bound to the encoded data for subsequent verification of data integrity. The final standardized control instruction data packet is stored in the format of "data packet ID-encoded data-verification code-generation timestamp". The data packet ID is generated using UUID (such as "123e4567-e89b-12d3-a456-426614174000") to ensure global uniqueness. The generation timestamp is accurate to milliseconds (such as "2025-06-11 09:30:00.123") to provide a time reference for instruction traceability.

[0080] The standardized control instruction data packets are broadcast to each energy equipment node through the blockchain network, triggering the automatic verification logic of the smart contract and generating instruction verification results. This step is crucial for ensuring the security and effectiveness of instruction transmission. It achieves decentralized broadcasting of instructions through the distributed network of blockchain, while automatically verifying the legality of instructions using the preset logic of smart contracts to prevent illegal instructions or tampering with instruction execution. The specific implementation method is as follows: The blockchain network adopts a consortium blockchain architecture (only authorized energy device nodes and control center nodes participate, balancing security and efficiency), with 20 nodes (covering all device clusters and the control center). It employs a Practical Byzantine Fault Tolerance (PBFT) consensus mechanism (consensus latency ≤500ms, meeting real-time control requirements). The broadcast process for standardized control command data packets is as follows: the control center node, as the initiator, signs the data packet (using the ECDSA encryption algorithm; the private key is exclusively kept by the control center, and the public key is deployed on the blockchain), and sends it to the three seed nodes of the consortium blockchain. After verifying the signature validity, the seed nodes simultaneously broadcast it to the remaining 16 device nodes. Each node returns a confirmation receipt upon receiving the packet, ensuring 100% broadcast coverage.

[0081] During broadcasting, data packets are transmitted via an encrypted channel (based on the TLS 1.3 protocol). Leveraging the immutability of blockchain, once a data packet is uploaded to the chain, all nodes store the same copy, preventing loss or tampering during transmission. For example, after the control center node initiates a broadcast, all device nodes receive the data packet within 500ms, and the hash value of the data packet stored on each node matches that of the initiator (both are "7a9f3d8c...2b7e"), confirming successful transmission.

[0082] The smart contract's automatic verification logic is pre-set on the blockchain. The trigger condition is "the device node receives the data packet and verifies the signature." The verification dimensions include four core aspects, and a comprehensive verification result is output after each aspect is verified. The first aspect is device identifier verification: The smart contract queries the authorized device list on the blockchain to confirm whether "device_id" exists and is in a normal state (e.g., not disabled). If the device ID "PV000099" is not in the authorized list, it is directly judged as "verification failed." The second aspect is parameter range verification: The constraints in "execution_constraints" are extracted, and the parameter values ​​of "operation_params" are checked to see if they are in compliance. For example, if the charging and discharging power of the energy storage battery instruction is "-45.0kW," and it exceeds the constraint "≥-40.0kW," it is judged as "verification failed." The third aspect is time period validity verification: It checks whether "control_period" is within a reasonable range of the current time (within the next 24 hours). If the instruction time period is "2025-06-10 09:00-10:00" (expired), it is judged as "verification failed." The fourth item is data integrity verification: calculate the SHA-256 hash value of the received data packet and compare it with the checksum included in the data packet. If they do not match (indicating that the data has been tampered with), it is determined as "verification failed".

[0083] The verification results are divided into five categories: "Verification Passed," "Verification Failed - Illegal Device," "Verification Failed - Parameter Exceeded Limits," "Verification Failed - Invalid Time Period," and "Verification Failed - Data Tampering." Each category generates a corresponding verification log (including the verification node ID, verification time, and reason for failure). For example, if the instruction for photovoltaic panel PV000001 is verified and the device ID is authorized, the parameters are within the constraints, the time period is valid, and the data is complete, a "Verification Passed" result is generated. However, if the instruction for charging pile EV000012 has a charging / discharging power of "65.0kW," which exceeds the constraint "≤60.0kW," a "Verification Failed - Parameter Exceeded Limits" result is generated, with the reason for failure being "output_power=65.0kW>60.0kW."

[0084] Based on the instruction verification results, the smart contract automatically executes the verified control instructions, issues specific control commands through the device control interface, and generates device control signals; This step is the core of instruction execution. The smart contract only executes verified instructions, transforming abstract parameters into physical control signals that the device can recognize through a standardized device control interface, ensuring the precise execution of control instructions. The specific implementation method is as follows: The smart contract execution logic is bound to the verification result, and execution is triggered only when the verification result is "verification passed". The execution process consists of three steps: "instruction parsing - interface adaptation - signal generation". First, instruction parsing is performed: the smart contract extracts "device_id", "control_period", and "operation_params" from the data packet and converts them into fields that the device control interface can recognize. For example, "output_power:48.0kW" is parsed as "power setting value: 48000W" (the unit is watts for easy device recognition), and "control_period:2025-06-1110:00-11:00" is parsed as "execution start time: 10:00:00, execution duration: 3600 seconds".

[0085] The equipment control interface adopts a standardized industrial interface protocol (compatible with the control requirements of mainstream energy equipment). Photovoltaic panels and wind turbines use the Modbus TCP protocol (transmission rate 100Mbps, latency ≤10ms), while energy storage batteries and charging piles use the OPCUA protocol (supporting complex parameter transmission). Interface communication uses CRC-32 verification to ensure error-free transmission of control commands. During interface adaptation, the smart contract automatically selects the corresponding protocol based on the device type of "device_id". For example, PV000001 represents a photovoltaic panel, so the Modbus TCP protocol is selected. The parsed parameters are encapsulated into a Protocol Data Unit (PDU), containing a function code (0x06 represents writing a single register), a register address (output power corresponds to register address 0x0001), and parameter values ​​(48000W converted to hexadecimal 0xBB80).

[0086] The generation of control signals must match the driving requirements of the equipment and is divided into two categories: analog signals and digital signals. Analog signals are used for continuous parameter control (such as output power adjustment) and adopt a 4-20mA current signal (4mA corresponds to 0W, 20mA corresponds to the rated power of the equipment, such as 20mA for a photovoltaic panel rated at 50kW and 19.2mA for 48kW). Digital signals are used for switching control (such as equipment start-up and shutdown) and adopt a 0-5V voltage signal (0V represents off, 5V represents on). For example, the control signal for the photovoltaic panel PV000001 is "4-20mA current signal, value 19.2mA, execution time 3600 seconds"; the discharge control signal for the energy storage battery BAT000003 is "4-20mA current signal, value 8mA (8mA for -30kW, 20mA for 40kW rated discharge), linked to the battery temperature monitoring signal (automatic interruption when the temperature exceeds 35℃)".

[0087] The control signal is received by the built-in signal receiving module of the device, and after digital-to-analog conversion (16-bit precision), it drives the device to execute. For example, after the MPPT controller of the photovoltaic panel receives a 19.2mA current signal, it adjusts the power module output to 48kW. After the charging pile receives the corresponding signal, it adjusts the output current and voltage of the charging module to ensure that the operating parameters are consistent with the instructions.

[0088] The system monitors the equipment's execution status in real time and records the results to a blockchain distributed ledger, ensuring the traceability and immutability of the control process, and ultimately achieving distributed collaborative optimization and safe management of regional energy equipment.

[0089] This step is the core of the closed-loop control system. It verifies the execution effect through real-time monitoring, achieves full-process traceability using blockchain ledgers, and optimizes collaborative strategies based on feedback data to ensure the stable operation of the regional energy system. The specific implementation method is as follows: Real-time monitoring of equipment performance is achieved through multimodal sensors deployed on the equipment. The sensor sampling frequency is synchronized with the control signal (10Hz for devices controlled by analog signals, and 1Hz for devices controlled by digital signals). Monitoring parameters include "actual operating parameters - equipment status - environmental parameters," such as "actual output power - operating status (operating / fault) - light intensity" for photovoltaic panels, and "actual charge / discharge power - battery SOC - battery temperature" for energy storage batteries. After preprocessing (outliers removed, units standardized) by edge computing nodes, the monitoring data is uploaded to the blockchain network in real time at a frequency of once per second to ensure real-time status feedback.

[0090] The execution results are recorded using a blockchain distributed ledger, with a ledger structure of "block header - block body". The block header contains the hash value of the previous block, the hash value of the current block, and a timestamp. The block body contains monitoring data, the instruction ID, and the execution result determination. The execution result determination is automatically generated by the smart contract based on the monitoring data and is divided into three categories: "execution successful", "partially successful", and "execution failed". When the deviation between the actual operating parameters and the instruction parameters is ≤ ±5%, it is determined as "execution successful" (e.g., instruction 48kW, actual 47.2kW, deviation 1.7%); when the deviation is between 5% and 10%, it is determined as "partially successful"; when the deviation is >10% or the equipment malfunctions, it is determined as "execution failed". For example, the instruction for energy storage battery BAT000003 to discharge 30kW was executed, but the actual discharge was 28.8kW, a deviation of 4%, which was judged as "execution successful". The execution result was recorded as "device_id:BAT000003,control_id:123e4567...,actual_power:-28.8kW,status:execution successful,timestamp:2025-06-1110:00:01.234".

[0091] The immutability of the distributed ledger is achieved through a hash chain. The hash value of each block is calculated from the hash value of the previous block and the data of the current block. If the data of a block is tampered with, its hash value will change, causing the hash chain of all subsequent blocks to break. All nodes can detect the anomaly, ensuring the authenticity of the records. Simultaneously, the ledger supports authorized queries. Authorized nodes (such as control centers and operations teams) can query execution records by device ID or time range, enabling full-chain traceability of the control process. For example, querying the execution record of PV000001 from 10:00 to 11:00 provides complete information such as instruction content, verification results, actual running curves, and status feedback.

[0092] Distributed collaborative optimization based on execution results is achieved through the feedback logic of smart contracts: If a device fails to execute (e.g., charging pile EV000012 cannot execute the 38kW charging command due to a fault), the smart contract automatically triggers a backup plan, adjusting the power allocation of other charging piles in the same cluster (e.g., distributing the 38kW of EV000012 to EV000013 and EV000014, increasing each by 19kW), and simultaneously broadcasting the fault information to the control center node to remind maintenance personnel to handle it. If multiple devices execute successfully and the system load is balanced (the deviation between the actual total load and the predicted load is ≤ ±3%), the subsequent commands remain unchanged; if the deviation is > 3%, the smart contract adjusts the command parameters for the next time period based on monitoring data (e.g., if the actual load is 5% higher than the prediction, increase the energy storage discharge power by 5kW), achieving dynamic collaborative optimization.

[0093] Security management is reflected in three levels: first, encrypted protection for command transmission and execution to prevent unauthorized intrusion; second, automatic verification and fault handling of smart contracts to prevent equipment overload or malfunction; and third, the traceability of the distributed ledger to facilitate fault tracing and liability determination. Ultimately, through the synergy of blockchain and smart contracts, decentralized, secure, and precise control of regional energy equipment can be achieved, realizing a multi-objective synergy of efficient renewable energy consumption, optimized operating costs, and stable system operation.

[0094] As can be seen, by deploying multimodal sensor arrays on energy equipment nodes to collect multi-source heterogeneous data including equipment operating status data and environmental parameters, standardized equipment operating feature vectors are generated. These feature vectors are then input into a regional energy load prediction model based on federated learning to generate regional load prediction curves for a preset future time period. Based on the regional load prediction curves and combined with equipment operating constraints, a multi-objective optimization algorithm is used to dynamically solve for the optimal operating parameters of each energy device, generating a set of equipment control commands. Based on this set of control commands, the control commands are automatically verified and executed through a blockchain-enabled smart contract execution layer. This achieves a unified approach to regional energy equipment in three dimensions: data privacy and security, precise and efficient control, and reliable command execution, thus achieving safe, economical, and self-consistent intelligent control.

[0095] Another embodiment of the present invention provides an intelligent control system for energy equipment based on the Internet of Things (IoT), see [link to relevant documentation]. Figure 3 The system may include: The acquisition module 301 is used to acquire multi-source heterogeneous data containing equipment operating status data and environmental parameters through a multi-modal sensor array deployed on energy equipment nodes, and to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data using edge computing nodes to generate a standardized equipment operating feature vector. Prediction module 302 is used to input the device operation feature vector into the regional energy load prediction model based on federated learning, and generate the regional load prediction curve for a future preset period by fusing the data features of multiple nodes without exchanging the original data. The generation module 303 is used to dynamically solve the optimal operating parameters of each energy device based on the partition load prediction curve and combined with the equipment operation constraints, and generate a set of equipment control instructions. The control module 304 is used to automatically verify and execute the control commands based on the set of device control commands through a blockchain-enabled smart contract execution layer, thereby realizing distributed collaborative optimization and safety management of regional energy equipment.

[0096] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0097] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0098] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0099] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. An Internet of Things-based intelligent regulation method for energy equipment, characterized in that, The method comprises: Collecting multi-source heterogeneous data containing equipment operation state data and environmental parameters through a multi-modal sensor array deployed at an energy equipment node, and performing spatio-temporal alignment and feature extraction on the multi-source heterogeneous data using an edge computing node to generate a standardized equipment operation feature vector; Inputting the equipment operation feature vector into a district energy load prediction model based on federated learning, generating a district load prediction curve for a future preset period by fusing multi-node data features without exchanging original data; According to the district load prediction curve, combining equipment operation constraint conditions, and dynamically solving optimal operation parameters of each energy equipment using a multi-objective optimization algorithm to generate a set of equipment control instructions; Based on the set of equipment control instructions, automatically verifying and executing the control instructions through a smart contract execution layer enabled by a blockchain to realize distributed collaborative optimization and safe management and control of district energy equipment.

2. The method of claim 1, wherein, The multi-modal sensor array deployed at the energy equipment node collects multi-source heterogeneous data containing equipment operation state data and environmental parameters, and the edge computing node performs spatio-temporal alignment and feature extraction on the multi-source heterogeneous data to generate a standardized equipment operation feature vector, comprising: Synchronously collecting equipment operation state data and environmental parameters through a multi-modal sensor array deployed at an energy equipment node to generate an original multi-source heterogeneous data set; Timestamp alignment and spatial position calibration are performed on the original multi-source heterogeneous data set, an interpolation algorithm is used to compensate for the sampling time difference of different sensors to generate a spatio-temporally aligned multi-source data sequence; Based on the spatio-temporally aligned multi-source data sequence, a feature extraction algorithm is used to extract key operation features to generate a multi-dimensional feature vector; The multi-dimensional feature vector is standardized by using Z-score normalization method to eliminate dimension difference, and finally a standardized equipment operation feature vector is generated.

3. The method of claim 2, wherein, The equipment operation feature vector is input into a district energy load prediction model based on federated learning to generate a district load prediction curve for a future preset period by fusing multi-node data features without exchanging original data, comprising: Initializing a federated learning client at each edge computing node, loading a pre-trained load prediction model, and inputting the standardized equipment operation feature vector into the local model for forward calculation to generate a local model update gradient; Using homomorphic encryption technology to encrypt the local model update gradient, transmitting the encrypted gradient to the federated learning server through a secure communication protocol to generate an encrypted gradient set; Performing secure aggregation calculation on the encrypted gradient set at the federated learning server side, fusing the model updates of each node using a federated averaging algorithm to generate global model update parameters; Updating the district energy load prediction model based on the global model update parameters, and predicting the load for a future preset period using the updated model to finally generate a district load prediction curve.

4. The method of claim 3, wherein, According to the district load prediction curve, combining equipment operation constraint conditions, and dynamically solving optimal operation parameters of each energy equipment using a multi-objective optimization algorithm to generate a set of equipment control instructions, comprising: The load distribution characteristics in the partition load prediction curve are analyzed, peak load period, load change trend and load space distribution information are extracted, and a load characteristic analysis report is generated; Based on the load characteristic analysis report, a multi-objective optimization model considering equipment operation constraints is established, the constraint conditions include equipment power upper and lower limits, operation efficiency interval and equipment start-stop limit, and an optimization problem mathematical model is generated; A multi-objective particle swarm optimization algorithm is used to solve the optimization problem mathematical model, and multiple objectives including energy utilization efficiency, equipment operation cost and system stability are optimized simultaneously to generate an optimal solution set; From the optimal solution set, select the Pareto optimal solution that meets the actual engineering demand, analyze it into specific equipment operation parameters, and finally generate a set of equipment control instructions.

5. The method of claim 4, wherein, Based on the set of equipment control instructions, the smart contract execution layer enabled by the blockchain automatically verifies and executes the control instructions to realize the distributed collaborative optimization and safe control of regional energy equipment, including: Encode the set of equipment control instructions into a data structure recognizable by the smart contract to generate a standardized control instruction data packet; Broadcast the standardized control instruction data packet to each energy equipment node through the blockchain network, trigger the automatic verification logic of the smart contract, and generate an instruction verification result; Based on the instruction verification result, the smart contract automatically executes the verified control instructions, issues specific control commands through the equipment control interface, and generates equipment control signals; Real-time monitoring of equipment execution status and recording of execution results to the distributed ledger of the blockchain ensures the traceability and tamper resistance of the control process, and finally realizes the distributed collaborative optimization and safe control of regional energy equipment.

6. An energy device intelligent regulation system based on an Internet of Things, characterized in that, The system comprises: The acquisition module is configured to acquire multi-source heterogeneous data including equipment operation state data and environmental parameters through a multi-modal sensor array deployed on an energy equipment node, and perform temporal and spatial alignment and feature extraction on the multi-source heterogeneous data using an edge computing node to generate a standardized equipment operation feature vector; The prediction module is configured to input the equipment operation feature vector into a regional energy load prediction model based on federated learning, and generate a partition load prediction curve for a future preset period by fusing multi-node data features without exchanging original data; The generation module is configured to dynamically solve the optimal operation parameters of each energy equipment using a multi-objective optimization algorithm based on the partition load prediction curve and the equipment operation constraint conditions to generate a set of equipment control instructions; The control module is configured to automatically verify and execute the control instructions based on the set of equipment control instructions through the smart contract execution layer enabled by the blockchain to realize the distributed collaborative optimization and safe control of regional energy equipment.

7. The system of claim 6, wherein, The acquisition module is specifically configured to: Synchronously acquire equipment operation state data and environmental parameters through a multi-modal sensor array deployed on an energy equipment node to generate an original multi-source heterogeneous data set; Timestamp alignment and spatial position calibration are performed on the original multi-source heterogeneous data set, and an interpolation algorithm is used to compensate for the sampling time difference of different sensors to generate a spatio-temporally aligned multi-source data sequence; Based on the spatio-temporally aligned multi-source data sequence, a feature extraction algorithm is used to extract key operation features to generate a multi-dimensional feature vector; The multi-dimensional feature vector is normalized, and a Z-score normalization method is used to eliminate dimension differences, and finally a standardized equipment operation feature vector is generated.

8. The system of claim 7, wherein, The prediction module is specifically configured to: Initialize a federated learning client at each edge computing node, load a pre-trained load prediction model, and input the standardized equipment operation feature vector into the local model for forward calculation to generate a local model update gradient; Perform encryption processing on the local model update gradient using homomorphic encryption technology, transmit the encrypted gradient to the federated learning server through a secure communication protocol, and generate an encrypted gradient set; Perform secure aggregation calculation on the encrypted gradient set at the federated learning server, fuse the model updates of each node using a federated averaging algorithm, and generate global model update parameters; Update the regional energy load prediction model based on the global model update parameters, and use the updated model to predict the load in a future preset time period to finally generate a partitioned load prediction curve.

9. A storage medium, characterized by The storage medium has a computer program stored therein, wherein the computer program is configured to execute the method of any one of claims 1-5 when running.

10. An electronic device comprising a memory and a processor, characterized in that, The memory has a computer program stored therein, and the processor is configured to execute the computer program to execute the method of any one of claims 1-5.

Images