Distributed power generation and supply system optimization method and system based on AI load prediction

Abstract

The invention relates to the technical field of artificial intelligence and power system crossing, and discloses a distributed power generation and power supply system optimization method and system based on AI load prediction, and the method comprises the steps: enabling each distributed energy node to collect local multi-source time series data, and carrying out the preprocessing of the local multi-source time series data; deploying a local AI prediction model fusing a gating circulation unit and a multi-head self-attention mechanism, and generating load-power generation combined prediction; the gradient is uploaded to a central coordination server through differential privacy disturbance and homomorphic encryption; the server issues global update quantity after safely aggregating the gradient, and model cooperative training under federated learning is realized; and outputting and generating a local scheduling strategy based on the updating model, and forming a collaborative optimization instruction after global power flow verification and conflict resolution. According to the system, on the premise that data privacy is guaranteed, the prediction precision, the economical efficiency and stability of power supply scheduling and the renewable energy consumption rate are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and power systems, specifically relating to an optimization method and system for distributed generation and power supply systems based on AI load forecasting. Background Technology

[0002] With the advancement of "dual-carbon" goals and the acceleration of energy structure transformation, distributed generation systems are widely used in industrial parks, commercial complexes, and community microgrids. These systems typically integrate renewable energy sources such as photovoltaics and wind power, and require real-time matching with local loads to improve energy efficiency and power supply reliability. Against this backdrop, AI-based load and power generation forecasting has become a core supporting technology for optimized scheduling, energy storage configuration, and grid interaction. However, the historical operational data accumulated by a single distributed unit is limited in scale and highly non-stationary in distribution, resulting in insufficient training samples for AI models, weak generalization ability, and predictive accuracy that fails to meet the control requirements of highly dynamic scenarios.

[0003] Cross-domain collaborative modeling based on federated learning offers a novel approach to solving the data silo problem. This method allows multiple participants to jointly train models by exchanging model parameter updates without sharing original sensitive data, thereby constructing a global prediction model with stronger generalization capabilities. For distributed generation systems, various parks or communities want to leverage external data to improve their AI prediction performance, but are unable to directly upload original load and generation records due to privacy concerns involving user electricity consumption behavior, production plans, and business operations.

[0004] While centralized cloud platforms aggregate local model parameters to improve prediction accuracy, the lack of effective measurement and incentive mechanisms for participant contributions can easily lead to "free-riding" behavior, weakening the alliance's willingness to collaborate in the long term. Furthermore, without credible verification and tamper-proof records during model updates, the fairness and auditability of the aggregation results are difficult to ensure. Especially in high-value application scenarios involving carbon trading, green electricity certification, or profit sharing, the lack of transparent and automated rights allocation mechanisms will severely restrict the sustainable development of cross-domain collaborative ecosystems. Therefore, how to build an optimized architecture for distributed power supply systems with credible contribution assessment, dynamic incentive feedback, and efficient model collaboration capabilities while ensuring privacy and security has become a pressing technical challenge. Summary of the Invention

[0005] This invention provides a method and system for optimizing distributed generation and power supply systems based on AI load forecasting. It aims to address the problem of insufficient accuracy in AI prediction models within a single park or community due to limited sample sizes of load and generation data. Simultaneously, it overcomes the technical obstacle of centralized data sharing across multiple nodes caused by user privacy and trade secret restrictions. This invention constructs a cross-domain collaborative modeling mechanism based on a federated learning architecture, combining localized feature extraction, encrypted gradient aggregation, and dynamic weight allocation strategies. Without exchanging original data, it achieves high-precision and robust joint load and generation forecasting, and generates globally optimized power dispatching instructions accordingly.

[0006] This invention provides an optimization method for distributed generation and power supply systems based on AI load forecasting, comprising: Local load and power generation data acquisition units are deployed at multiple distributed energy nodes to acquire historical load power sequences, distributed power output sequences, meteorological and environmental parameter sequences, and equipment operation status sequences of each node in real time. The sequence data are preprocessed locally to form a standardized time-series feature tensor; a local AI prediction model is deployed at each distributed energy node. The local AI prediction model is a neural network structure based on the fusion of gated recurrent units and attention mechanisms, which is used to generate local load-generation joint prediction results based on the standardized time-series feature tensor. Local model training is performed at each distributed energy node, and the gradient vector of the local loss function with respect to the model parameters is calculated. The gradient vector is then processed with differential privacy perturbation and homomorphic encryption before being uploaded to the central coordination server. The central coordination server receives encrypted gradient vectors uploaded by all distributed energy nodes, performs secure aggregation operations, and generates global gradient update quantities. The central coordination server broadcasts the global gradient update to all distributed energy nodes; Each distributed energy node updates the parameters of its local AI prediction model using the global gradient update amount; based on the future load and power generation prediction values ​​output by the updated local AI prediction model, combined with grid dispatch constraints and economic objective function, a local optimal power supply dispatch strategy is generated. The local optimal power supply scheduling strategy is uploaded to the central coordination server; the central coordination server aggregates the scheduling strategies of all nodes, performs global power flow verification and conflict resolution, generates the final collaborative optimization scheduling instruction, and issues it to each distributed energy node for execution.

[0007] In one embodiment of the present invention, the historical load power sequence has a time resolution of 15 minutes and a sequence length of 720 time points, corresponding to the complete load curve of the past 7 days; the distributed power output sequence includes the output power of the photovoltaic array and the output power of the small wind turbine, and the sampling frequency is synchronized with the load power sequence; the meteorological environment parameter sequence includes solar irradiance, ambient temperature, relative humidity, wind speed and wind direction, which are collected by micro weather stations deployed at each node at 10-minute intervals and aligned to a 15-minute time grid through linear interpolation; the equipment operation status sequence includes inverter operating mode, energy storage battery state of charge, switching device on / off status and fault alarm flags, which are read in real time through fieldbus protocol.

[0008] As one embodiment of the present invention, the local preprocessing includes missing value detection and linear imputation, outlier identification and sliding window midpoint filtering, and normalization processing; the normalization processing uses the maximum and minimum values ​​of each sequence in the most recent 30 days for linear scaling, mapping all feature dimensions to the interval between 0 and 1.

[0009] In one embodiment of the present invention, the input layer of the local AI prediction model receives a four-dimensional feature tensor, the dimensions of which are time step, number of feature channels, spatial location index, and auxiliary context identifier; the number of units in the hidden layer of the gated recurrent unit is 256, and the number of stacked layers is three; the attention mechanism is a multi-head self-attention structure with 8 heads, used to dynamically weight the contribution of different historical time steps to the current prediction moment; the model output layer generates load power prediction values ​​and power generation prediction values ​​for the next 96 time steps, with a time step of 15 minutes, covering the next 24 hours.

[0010] In one embodiment of the present invention, the local loss function is a weighted mean square error function, and its weight coefficients are dynamically adjusted according to the distance of the prediction time step, with lower weights for more distant time steps; specifically, the first... The weights of each prediction time step are: ,in From 1 to 96.

[0011] In one embodiment of the present invention, the differential privacy perturbation is applied before the gradient vector is uploaded, the perturbation noise follows a Laplace distribution, and the scale parameter is 0.05; the homomorphic encryption adopts a lattice-based fully homomorphic encryption scheme, which supports the performance of addition and scalar multiplication operations in the ciphertext state.

[0012] As one embodiment of the present invention, the secure aggregation operation includes: the central coordination server first verifies the integrity and source legitimacy of the gradient vectors uploaded by each node; then, it performs element-wise summation on all valid gradient vectors and divides them by the total number of nodes participating in the aggregation to obtain the unencrypted global gradient average value; this average value is then decrypted by the private key of the central coordination server and broadcast as the global gradient update.

[0013] As one embodiment of the present invention, the power grid dispatch constraints include upper and lower limits of node voltage amplitude, upper limit of line transmission capacity, limit of distributed power source ramp rate, limit of energy storage system charge and discharge depth, and minimum start-up and shutdown time constraints; the economic objective function is to minimize operating costs, including electricity purchase costs, fuel consumption costs, equipment depreciation and loss, and carbon emission penalties.

[0014] In one embodiment of the present invention, the global power flow verification employs the Newton-Raphson method to iteratively solve the power flow equations of the distribution network, with the convergence criterion being that the node power imbalance is less than [a certain value]. If there is a situation where the power flow does not converge or exceeds the limit, the conflict resolution mechanism will be activated. This mechanism will prioritize reducing the scheduling instruction deviation of the node with the largest prediction error until the network-wide operation constraints are met.

[0015] This invention provides an optimization system for distributed generation and power supply systems based on AI load forecasting, comprising: The distributed data acquisition subsystem is deployed at each distributed energy node to acquire historical load power sequences, distributed power output sequences, meteorological and environmental parameter sequences, and equipment operating status sequences in real time. The local data preprocessing module is used to perform missing value imputation, outlier filtering and normalization on the collected data to generate standardized time series feature tensors; The local AI prediction model module is a neural network that integrates a gated recurrent unit and a multi-head self-attention mechanism, used to generate joint forecasts of load and power generation for future periods. The local model training and gradient generation module is used to calculate the gradient vector of the local loss function and apply differential privacy perturbation and homomorphic encryption; the central coordination server includes an encrypted gradient receiving unit, a secure aggregation unit, a global model update broadcasting unit, a scheduling strategy aggregation unit, a global power flow verification unit, and a conflict resolution unit. The local power supply scheduling strategy generation module is used to generate the local optimal scheduling strategy based on the updated local AI prediction model output, combined with operational constraints and economic objectives. The collaborative optimization scheduling instruction distribution module is used to distribute the final instructions generated by the central coordination server to each node for execution.

[0016] As one embodiment of the present invention, the distributed data acquisition subsystem consists of a smart meter, a photovoltaic inverter built-in monitoring unit, a wind turbine controller, a micro weather station, and an equipment status monitoring terminal. Each unit is connected to the local edge computing gateway via industrial Ethernet or power line carrier communication.

[0017] As one embodiment of the present invention, the local edge computing gateway is equipped with a dual-core processor with a main frequency of 1.2 GHz, a memory capacity of 4 gigabytes, a storage medium of 32 gigabyte solid-state drive, runs an embedded real-time operating system, and supports containerized deployment of local AI prediction model modules.

[0018] As one embodiment of the present invention, the central coordination server is equipped with a 16-core central processing unit, 64 gigabytes of memory and a 1 terabyte solid-state storage array, runs a distributed task scheduling middleware, and supports real-time response capability with each aggregation cycle not exceeding 5 minutes.

[0019] As one embodiment of the present invention, the training cycle of the local AI prediction model module is once a day, which is fixed at 2:00 AM; each training uses local data from the past 7 days as the training set, adopts a stochastic gradient descent optimizer, and has an initial learning rate of 0.001 and a momentum coefficient of 0.9.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by introducing a federated learning framework, enables multiple distributed energy nodes to collaboratively train a high-precision AI prediction model without leaving the domain of the original data, effectively overcoming the problem of insufficient model generalization ability caused by the sparsity of single-point data. The dual protection mechanism of differential privacy and homomorphic encryption ensures the security of gradient information during transmission and aggregation, meeting compliance requirements for user privacy and trade secrets.

[0021] 2. The local model employs a gated cyclic unit and a multi-head self-attention fusion structure, which can accurately capture long-term dependencies and nonlinear abrupt changes at key time points in the load and generation sequences. The secure aggregation and global power flow verification mechanisms executed by the central coordination server ensure the consistency of model updates and the grid feasibility of scheduling commands.

[0022] 3. Without sacrificing data sovereignty, this invention significantly improves the accuracy of load and power generation forecasts, thereby optimizing the operational economy, stability, and renewable energy absorption rate of distributed power supply systems. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall technical architecture of the distributed generation power supply system optimization method and system based on AI load forecasting proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the cross-domain collaborative AI prediction model based on federated learning in this invention; Figure 3 This is a logical flow diagram of the local data acquisition and preprocessing stage in this invention; Figure 4 This is a logical flow diagram of the local AI prediction model training and encrypted gradient uploading stage in this invention; Figure 5 This is a logical flow diagram of the central coordination server performing secure aggregation and global model updates in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between local scheduling strategy generation and global collaborative optimization scheduling instruction issuance in this invention. Detailed Implementation

[0024] Please refer to the attached document. Figures 1 to 6 This invention provides a method and system for optimizing distributed generation and power supply systems based on artificial intelligence load forecasting. Its core lies in constructing a federated learning-driven cross-domain collaborative modeling mechanism. This mechanism achieves high-precision and robust joint load and generation forecasting while ensuring that the original data of each distributed energy node does not leave its local area. Based on this forecast, collaborative optimization scheduling instructions that meet grid operation constraints and economic objectives are generated. The specific embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0025] The method includes the following steps: Local load and power generation data acquisition units are deployed at multiple distributed energy nodes to acquire historical load power sequences, distributed power output sequences, meteorological and environmental parameter sequences, and equipment operation status sequences of each node in real time. The sequence data are preprocessed locally to form a standardized time-series feature tensor; a local artificial intelligence prediction model is deployed at each distributed energy node. The local artificial intelligence prediction model is a neural network structure based on the fusion of gated recurrent units and attention mechanism, which is used to generate local load-generation joint prediction results based on the standardized time-series feature tensor; local model training is performed at each distributed energy node, and the gradient vector of the local loss function with respect to the model parameters is calculated. The gradient vector is processed by differential privacy perturbation and homomorphic encryption and then uploaded to the central coordination server. The central coordination server receives the encrypted gradient vectors uploaded by all distributed energy nodes, performs a secure aggregation operation, and generates a global gradient update. The central coordination server broadcasts the global gradient update to all distributed energy nodes; each distributed energy node uses the global gradient update to update the parameters of its local artificial intelligence prediction model; based on the future load and power generation prediction values ​​output by the updated local artificial intelligence prediction model, combined with grid dispatch constraints and economic objective function, a local optimal power supply dispatch strategy is generated. The local optimal power supply scheduling strategy is uploaded to the central coordination server; the central coordination server aggregates the scheduling strategies of all nodes, performs global power flow verification and conflict resolution, generates the final collaborative optimization scheduling instruction, and issues it to each distributed energy node for execution.

[0026] In one embodiment of the present invention, the historical load power sequence has a time resolution of 15 minutes and a sequence length of 720 time points, corresponding to the complete load curve of the past 7 days. This time resolution is sufficient to capture typical electricity consumption patterns within the park or community, such as weekday morning and evening peak hours and weekend off-peak hours. The distributed power output sequence includes the output power of the photovoltaic array and the output power of the small wind turbine, with the sampling frequency synchronized with the load power sequence to ensure time alignment of the three. The meteorological environmental parameter sequence includes solar irradiance, ambient temperature, relative humidity, wind speed, and wind direction, collected by micro-weather stations deployed at each node at 10-minute intervals, and aligned to a 15-minute time grid through linear interpolation to match the sampling rhythm of other sequences. The equipment operation status sequence includes inverter operating mode, energy storage battery state of charge, switching device on / off status, and fault alarm flags, read in real time via fieldbus protocol with a sampling period of 1 second, and subsequently downsampled to a 15-minute granularity through moving average to maintain overall data consistency.

[0027] In one embodiment of the present invention, the local preprocessing includes missing value detection and linear imputation, outlier identification and sliding window median filtering, and normalization. Missing value detection identifies data points as missing if there are more than three consecutive zero values, and imputs them using linear interpolation of adjacent non-missing points. Outlier identification is based on a 3-standard-deviation criterion; data points exceeding the mean plus or minus 3 standard deviations are marked as outliers and replaced with a median filter with a window size of 9 time steps. The normalization process uses linear scaling of the maximum and minimum values ​​of each sequence within the most recent 30 days, mapping all feature dimensions to the 0-1 interval, as expressed by the formula: ; in, For the original data points, and These are the minimum and maximum values ​​of this feature over the past 30 days, respectively. This is the normalized value. This normalization strategy avoids model training bias caused by differences in the dimensions of different features, while preserving local dynamic range information.

[0028] In one embodiment of the present invention, the input layer of the local artificial intelligence prediction model receives a four-dimensional feature tensor, the dimensions of which are time step, number of feature channels, spatial location index, and auxiliary context identifier. The time step dimension is fixed at 720, corresponding to 7 days of historical data; the number of feature channels is 12, including load power, photovoltaic output, wind power output, solar irradiance, ambient temperature, relative humidity, wind speed, wind direction, inverter mode, battery state of charge, switch status, and fault flag; the spatial location index is used to distinguish measurement points in multiple sub-areas within the same park, and is always 1 in a single-node scenario; the auxiliary context identifier encodes the current date type (weekday, weekend, holiday) and seasonal information to enhance the model's ability to perceive periodic patterns.

[0029] The gated recurrent unit has 256 hidden layers and is stacked in three layers. The output of each layer is passed to the next layer via a residual connection to alleviate the gradient vanishing problem. The attention mechanism is a multi-head self-attention structure with 8 heads, used to dynamically weight the contribution of different historical time steps to the current prediction time. Specifically, the... The query, key, and value matrix of each attention head is obtained by linear transformation of the hidden state of the last layer of the gated recurrent unit. The attention weight is calculated using the following formula: ; in, For the first The value matrix of each attention head, For the first The key matrix of each attention head. For the first A query matrix with attention heads The dimension of the key vector is 32. The outputs of the 8 heads are concatenated, then dimensionality reduced by a fully connected layer, and then fused with the output of the gated recurrent unit before being fed into the decoder. The model output layer generates load power predictions and power generation predictions for the next 96 time steps, with a time step of 15 minutes, covering the next 24 hours, and an output dimension of 192 (96 time steps × 2 prediction targets).

[0030] In one embodiment of the present invention, the local loss function is a weighted mean square error function, whose weight coefficients are dynamically adjusted according to the distance of the prediction time step, with lower weights for more distant time steps. Specifically, the first... The weights of each prediction time step are: ,in The weights range from 1 to 96. This weight design allows the model to focus more on recent prediction accuracy during training, which aligns with the higher requirements of practical scheduling for short-term prediction reliability. The loss function expression is: ; in, and These are the actual load and the generated power, respectively. and These are the model's predicted values.

[0031] In one embodiment of the present invention, the differential privacy perturbation is applied before the gradient vector is uploaded, and the perturbation noise follows a Laplace distribution with a scale parameter of 0.05. This scale parameter has been experimentally verified to provide [property / privacy] while maintaining a model utility loss of less than 2%. Differential privacy protection is provided. The homomorphic encryption adopts a lattice-based fully homomorphic encryption scheme, which supports addition and scalar multiplication operations in the ciphertext state. In the specific implementation, each node uses the public key distributed by the central coordination server to encrypt the perturbed gradient vector element by element, generating a ciphertext gradient vector, ensuring that even if the communication link is eavesdropped, attackers cannot restore the original gradient value or infer the local data distribution.

[0032] As one embodiment of the present invention, the secure aggregation operation includes: the central coordination server first verifies the integrity and legality of the gradient vectors uploaded by each node, and confirms that the data has not been tampered with and originates from the registered node through a dual verification mechanism of digital signature and message authentication code; subsequently, it performs element-wise summation on all valid gradient vectors and divides by the total number of nodes participating in the aggregation. The unencrypted global gradient average is obtained; this average is then decrypted using the private key of the central coordinating server and broadcast as the global gradient update. The aggregation process satisfies the security requirements of secure multi-party computation; any value less than [a certain value] will be excluded. The collusion of a few nodes makes it impossible to infer the gradient information of the remaining nodes.

[0033] As one embodiment of the present invention, the power grid dispatch constraints include upper and lower limits of node voltage amplitude (per unit value 0.95 to 1.05), upper limit of line transmission capacity (set according to conductor thermal stability limit), distributed power source ramp rate limit (photovoltaic power changes no more than 10% of rated power every 15 minutes, wind power no more than 15%), energy storage system charge and discharge depth limit (state of charge maintained between 20% and 90%), and minimum start-stop time constraint (diesel generator start-stop interval no less than two hours). The economic objective function is to minimize operating costs, including electricity purchase costs. (Billed at time of day) Fuel consumption cost (For standby diesel generators) Equipment depreciation and loss (Related to the number of charge-discharge cycles) and carbon emission penalties (Converted based on local carbon trading prices). The optimal local power dispatch strategy is obtained by solving the following optimization problem: ; ; in, The vector of scheduling decision variables includes the output of each controllable power source, the charging and discharging power of energy storage, and the power purchase and sale plan, etc. This is the feasible region defined by the above constraints.

[0034] In one embodiment of the present invention, the global power flow verification employs the Newton-Raphson method to iteratively solve the power flow equations of the distribution network, with the convergence criterion being that the node power imbalance is less than [a certain value]. If power flow fails to converge or exceeds limits, a conflict resolution mechanism is activated. This mechanism first calculates the absolute value of the deviation between the scheduling instructions and the predicted values ​​of each node, and sorts them in descending order. It then prioritizes reducing the scheduling instruction deviation of the node with the largest prediction error, reducing it by 10% of its deviation each time, until the power flow of the entire network converges and all operating variables satisfy the constraints. This strategy ensures that the system can still generate physically feasible scheduling schemes even when there are local deviations in the model prediction.

[0035] In one embodiment of the present invention, the system includes a distributed data acquisition subsystem, a local data preprocessing module, a local artificial intelligence prediction model module, a local model training and gradient generation module, a central coordination server, a local power supply scheduling strategy generation module, and a collaborative optimization scheduling instruction issuance module. The distributed data acquisition subsystem consists of smart meters, a photovoltaic inverter built-in monitoring unit, a wind turbine controller, a micro weather station, and an equipment status monitoring terminal. Each unit is connected to a local edge computing gateway via industrial Ethernet or power line carrier communication. The local edge computing gateway is equipped with a dual-core processor with a clock speed of 1.2 GHz, 4 gigabytes of memory, and a 32-gigabyte solid-state drive. It runs an embedded real-time operating system and supports containerized deployment of the local artificial intelligence prediction model module. The central coordination server is equipped with a 16-core CPU, 64 gigabytes of memory, and a 1-terabyte solid-state storage array. It runs a distributed task scheduling middleware and supports a real-time response capability of no more than 5 minutes per aggregation cycle. The training cycle of the local artificial intelligence prediction model module is once a day, which is fixed at 2:00 AM. Each training uses local data from the past 7 days as the training set, adopts a stochastic gradient descent optimizer, has an initial learning rate of 0.001, a momentum coefficient of 0.9, and a fixed number of training rounds of 50.

[0036] In actual operation, each distributed energy node initiates its local training process at 2:00 AM daily: First, it loads the standardized time-series feature tensors of the past 7 days from its local database and inputs them into its local AI prediction model; it generates load and generation forecasts for the next 24 hours through forward propagation; it calculates the weighted mean square error loss; it performs backpropagation to obtain the gradient vector of the model parameters; it applies Laplace noise perturbation to this gradient vector and performs fully homomorphic encryption using the public key of the central coordination server; and it uploads the encrypted gradient to the central coordination server. Upon receiving the encrypted gradients from all registered nodes, the central coordination server performs secure aggregation, generates a global gradient update, and broadcasts it. Each node receives this update and uses it to adjust its local model parameters. Subsequently, based on the updated model, each node generates its local optimal power supply scheduling strategy for the next 24 hours and uploads it to the central coordination server. The central coordination server aggregates all strategies and performs a global power flow verification. If the verification passes, the strategy is directly implemented; if there are any violations, conflict resolution is initiated, a corrected collaborative optimization scheduling instruction is generated, and it is then implemented. The entire process is completed within 4 hours, ensuring that dispatch instructions take effect before 6 a.m. on the same day, covering peak electricity consumption throughout the day.

[0037] This embodiment achieves model co-evolution under data privacy protection through a federated learning architecture, improves the accuracy of time series prediction through gated recurrent units and multi-head self-attention fusion structure, and ensures the feasibility and security of scheduling instructions through secure aggregation and global power flow verification. Thus, without centralizing the original data, it significantly improves the prediction accuracy, operating economy and renewable energy absorption capacity of the distributed generation and power supply system.

Claims

1. An optimization method for distributed generation and power supply systems based on AI load forecasting, characterized in that, include: Local load and power generation data acquisition units are deployed at multiple distributed energy nodes to acquire historical load power sequences, distributed power output sequences, meteorological and environmental parameter sequences, and equipment operation status sequences of each node in real time. The sequence data are preprocessed locally to form a standardized temporal feature tensor; A local AI prediction model is deployed at each distributed energy node. The local AI prediction model is a neural network structure based on the fusion of gated recurrent units and attention mechanisms, which is used to generate local load-generation joint prediction results based on the standardized time-series feature tensor. Local model training is performed at each distributed energy node, and the gradient vector of the local loss function with respect to the model parameters is calculated. The gradient vector is then processed with differential privacy perturbation and homomorphic encryption before being uploaded to the central coordination server. The central coordination server receives encrypted gradient vectors uploaded by all distributed energy nodes, performs secure aggregation operations, and generates global gradient update quantities. The central coordination server broadcasts the global gradient update to all distributed energy nodes; Each distributed energy node updates the parameters of its local AI prediction model using the global gradient update amount; Based on the updated local AI prediction model outputting future load and generation forecasts, combined with grid dispatch constraints and economic objective functions, a local optimal power supply dispatch strategy is generated. The local optimal power supply scheduling strategy is uploaded to the central coordination server. The central coordination server aggregates the scheduling strategies of all nodes, performs global power flow verification and conflict resolution, generates the final collaborative optimization scheduling instructions, and issues them to each distributed energy node for execution.

2. The method for optimizing distributed generation and power supply systems based on AI load forecasting according to claim 1, characterized in that, The historical load power sequence has a time resolution of 15 minutes and a sequence length of 720 time points; The distributed power output sequence includes the output power of the photovoltaic array and the output power of the small wind turbine. The meteorological environmental parameter sequence includes solar irradiance, ambient temperature, relative humidity, wind speed and wind direction, which are collected by micro weather stations at 10-minute intervals and aligned to a 15-minute time grid through linear interpolation; The equipment operating status sequence includes inverter operating mode, energy storage battery state of charge, switching device on / off status, and fault alarm flag.

3. The method for optimizing distributed generation and power supply systems based on AI load forecasting according to claim 2, characterized in that, The local preprocessing includes missing value detection and linear imputation, outlier identification and sliding window mid-value filtering, and normalization processing; The normalization process uses the maximum and minimum values ​​of each sequence in the most recent 30 days to perform linear scaling, mapping all feature dimensions to the range of 0 to 1.

4. The method for optimizing distributed generation and power supply systems based on AI load forecasting according to claim 3, characterized in that, The input layer of the local AI prediction model receives a four-dimensional feature tensor, whose dimensions are time step, number of feature channels, spatial location index, and auxiliary context identifier. The gated loop unit has 256 units in its hidden layer and three stacked layers. The attention mechanism is a multi-head self-attention structure with 8 heads. The model output layer generates load power predictions and power generation predictions for the next 96 time steps, with a time step of 15 minutes.

5. The method for optimizing distributed generation and power supply systems based on AI load forecasting according to claim 4, characterized in that, The local loss function is a weighted mean square error function, and its weight coefficients are dynamically adjusted according to the distance of the prediction time step. No. The weights of each prediction time step are: ,in From 1 to 96.

6. The method for optimizing a distributed generation and power supply system based on AI load forecasting according to claim 5, characterized in that, The differential privacy perturbation is applied before the gradient vector is uploaded, and the perturbation noise follows a Laplace distribution with a scale parameter of 0.

05. The homomorphic encryption adopts a lattice-based fully homomorphic encryption scheme, which supports addition and scalar multiplication operations in the ciphertext state.

7. The method for optimizing distributed generation and power supply systems based on AI load forecasting according to claim 6, characterized in that, The secure aggregation operation includes: The central coordination server verifies the integrity and source legitimacy of the gradient vectors uploaded by each node; it performs element-wise summation on all valid gradient vectors and divides them by the total number of nodes participating in the aggregation to obtain the unencrypted global gradient average. This average value, after being decrypted by the private key of the central coordination server, is broadcast as a global gradient update.

8. The method for optimizing distributed generation and power supply systems based on AI load forecasting according to claim 7, characterized in that, The grid dispatch constraints include upper and lower limits of node voltage amplitude, upper limit of line transmission capacity, limit of distributed power source ramp rate, limit of energy storage system charge and discharge depth, and minimum start-up and shutdown time constraints; the economic objective function is to minimize operating costs, including electricity purchase costs, fuel consumption costs, equipment depreciation and loss, and carbon emission penalties.

9. The method for optimizing a distributed generation and power supply system based on AI load forecasting according to claim 8, characterized in that, The global power flow verification employs the Newton-Raphson method to iteratively solve the distribution network power flow equations. The convergence criterion is that the node power imbalance is less than [a certain value]. If there is a situation where the power flow does not converge or exceeds the limit, the conflict resolution mechanism will be activated to prioritize reducing the scheduling instruction deviation of the node with the largest prediction error until the network-wide operation constraints are met.

10. An optimization system for distributed generation and power supply systems based on AI load forecasting, characterized in that, include: The distributed data acquisition subsystem is used to acquire historical load power sequences, distributed power output sequences, meteorological and environmental parameter sequences, and equipment operating status sequences in real time from multiple distributed energy nodes. The local data preprocessing module is used to perform missing value imputation, outlier filtering and normalization on the sequence data to form a standardized time series feature tensor; The local AI prediction model module is a neural network that integrates a gated recurrent unit and a multi-head self-attention mechanism, used to generate joint forecasts of load and power generation for future periods. The local model training and gradient generation module is used to calculate the gradient vector of the local loss function and apply differential privacy perturbation and homomorphic encryption. The central coordination server includes an encrypted gradient receiving unit, a secure aggregation unit, a global model update broadcasting unit, a scheduling strategy aggregation unit, a global power flow verification unit, and a conflict resolution unit. The local power supply dispatching strategy generation module is used to generate the local optimal power supply dispatching strategy based on the updated local AI prediction model output, combined with grid dispatching constraints and economic objective function. The collaborative optimization scheduling instruction distribution module is used to distribute the final collaborative optimization scheduling instructions generated by the central coordination server to each distributed energy node for execution.

Images