Internet-based artificial intelligence power dispatching management system

Abstract

The invention discloses an Internet-based artificial intelligence power dispatching management system, which relates to the technical field of power system intelligent dispatching and comprises a distributed data acquisition module, a deep neural network module, a hierarchical reinforcement learning module, an intelligent execution module and a safety management module. The distributed data acquisition module acquires operation data of the power system through a multi-dimensional sensor array to form a feature data set; the deep neural network module processes the feature data set based on a double-layer cross validation mechanism to generate state evaluation data; the hierarchical reinforcement learning module constructs a multi-target power dispatching model and dynamically generates a control parameter matrix; and the intelligent execution module realizes parameter self-tuning through a built-in self-correction controller. According to the method, an improved Transform structure is designed, deep extraction is performed on the characteristics of the power system by fusing the space attention network and the time sequence attention network, and the model performance is optimized by adopting a three-stage strategy, so that the accuracy of state evaluation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent dispatching of electric power systems, and in particular to an artificial intelligence electric power dispatching management system and system based on the Internet. Background Art

[0002] The current power system dispatch management mainly adopts the traditional manual experience decision-making model, and the dispatchers formulate dispatch plans based on experience judgment and fixed rules. This method is difficult to respond to load fluctuations in a timely manner when facing a complex and changeable power grid operating environment. The real-time and accuracy of dispatch decisions are insufficient, which easily causes power supply quality fluctuations and economic benefits losses.

[0003] The existing power dispatching and management system also has obvious deficiencies in data collection, analysis and control. On the one hand, the traditional data collection equipment has limited accuracy and poor sampling synchronization, making it difficult to accurately capture the transient characteristics of power grid operation; on the other hand, the system lacks intelligent data analysis and adaptive adjustment capabilities, and the control strategy is relatively rigid, unable to be optimized and adjusted in real time according to the power grid status, making it difficult to achieve a good balance between power supply stability and economy. Summary of the invention

[0004] In view of the deficiencies in the existing power dispatching and management systems in terms of intelligence, control accuracy and self-adaptability, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to achieve high-precision collection of power grid operation data, intelligent evaluation of status, and adaptive optimization of dispatching strategies.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In the first aspect, an embodiment of the present invention provides an artificial intelligence power dispatching and management system based on the Internet, which includes a distributed data acquisition module, a deep neural network module, a hierarchical reinforcement learning module, an intelligent execution module and a safety management module, wherein: the distributed data acquisition module obtains power unit operation data, power grid transmission and distribution load data and environmental monitoring data through a multidimensional sensor array to form a feature data set; the multidimensional sensor array integrates a temperature sensing circuit, a voltage and current sampling circuit and a meteorological sensing circuit; the deep neural network module processes the feature data set based on a double-layer cross-validation mechanism to generate state evaluation data; the double-layer cross-validation mechanism includes a time domain feature extraction layer and a frequency domain feature analysis layer; the deep neural network module applies a Transformer structure, and the Transformer structure integrates a spatial attention network and a temporal attention network; the hierarchical reinforcement learning module constructs a multi-objective power dispatching model based on the state evaluation data; the hierarchical reinforcement learning module configures a dynamic weight update circuit and generates a control parameter matrix based on an online learning algorithm; the intelligent execution module receives the control signal and control parameter matrix output by the multi-objective power dispatching model, and the intelligent execution module has a built-in self-correction controller, and the self-correction controller performs parameter self-tuning according to the control parameter matrix; the safety management module performs safety monitoring on the entire system operation process.

[0007] As a preferred solution of the Internet-based artificial intelligence power dispatching and management system described in the present invention, wherein: the distributed data acquisition module configures the edge computing architecture to preprocess the feature data set; the double-layer cross-validation mechanism sets the data screening threshold to filter out abnormal data in the feature data set; the self-correcting controller is provided with a feedback compensation circuit to correct the control deviation in real time.

[0008] As a preferred solution of the Internet-based artificial intelligence power dispatching and management system described in the present invention, the deep neural network module includes the following steps: the time domain feature extraction layer performs segmentation processing on the feature data set to generate time domain statistical features; the frequency domain feature analysis layer performs wavelet transform on the time domain statistical features to obtain frequency domain energy distribution features; the spatial attention network fuses the time domain statistical features and the frequency domain energy distribution features, calculates the correlation weights of different monitoring points, and obtains the spatial feature vector; the temporal attention network extracts the time series pattern of the spatial feature vector to obtain the key temporal features; the Transformer structure inputs the spatial feature vector and the key temporal features into the multi-head self-attention layer and the feedforward neural network to generate state evaluation data.

[0009] As a preferred solution of the Internet-based artificial intelligence power dispatching and management system described in the present invention, the hierarchical reinforcement learning module includes: hierarchically extracting power load characteristics, power supply stability characteristics and economic cost characteristics in state evaluation data based on a sampling strategy to construct a feature vector space; setting a hierarchical scheduling engine, the hierarchical scheduling engine includes a day-ahead scheduling sublayer and a real-time scheduling sublayer; the day-ahead scheduling sublayer adopts a dual DQN network to map the power supply stability characteristics and economic cost characteristics into a power supply stability constraint function and an operating economy constraint function respectively; the real-time scheduling sublayer performs load prediction based on an LSTM model to construct a dual-objective optimization model; the dynamic weight update circuit adjusts the weight parameters of power supply stability and economy in the dual-objective optimization model according to the real-time load state and the prediction deviation, and updates the model parameters based on the gradient descent algorithm and the back propagation algorithm to obtain a control parameter matrix; the control parameter matrix includes a control gain coefficient, an integral time constant and a differential time constant.

[0010] As a preferred solution of the Internet-based artificial intelligence power dispatching and management system described in the present invention, the self-correcting controller performs parameter self-tuning according to the control parameter matrix, including: the self-correcting controller establishes a parameter mapping relationship based on a fuzzy reasoning mechanism, and the parameter mapping relationship includes a control gain factor and a time constant factor; the self-correcting controller optimizes and adjusts the control signal according to the parameter mapping relationship; the self-correcting controller uses an incremental tuning method to adaptively adjust the control gain factor and the time constant factor; the self-correcting controller generates a corrected control instruction based on the incremental tuning method.

[0011] As a preferred solution of the Internet-based artificial intelligence power dispatching and management system described in the present invention, the intelligent execution module includes: setting a control signal processing unit, the control signal processing unit receives the control signal output by the multi-objective power dispatching model, and performs signal analysis and classification storage; setting a parameter matrix analysis unit, the parameter matrix analysis unit performs feature decomposition on the control parameter matrix to obtain the control gain coefficient, integral time constant and differential time constant; the self-correcting controller establishes a control parameter mapping rule base based on the fuzzy reasoning mechanism, and the control parameter mapping rule base includes the control error magnitude and the error change rate; the self-correcting controller calculates the control parameter correction amount according to the control error magnitude and the error change rate, and the control parameter correction amount is used to dynamically adjust the control gain coefficient, integral time constant and differential time constant; the self-correcting controller sets a parameter setting module, and the parameter setting module is based on the adaptive PID control algorithm, and the control parameter is dynamically adjusted according to the control parameter correction amount to complete parameter self-tuning.

[0012] As a preferred solution of the Internet-based artificial intelligence power dispatching and management system described in the present invention, the distributed data acquisition module is provided with a data quality assessment unit, which uses a multi-level data screening algorithm to perform data integrity check and validity verification on the feature data set; the deep neural network module is provided with a model evaluation unit, which performs online evaluation on the accuracy of the state evaluation data based on the cross entropy loss function.

[0013] In the second aspect, an embodiment of the present invention provides an Internet-based artificial intelligence power dispatching management method, which includes a distributed data acquisition module that acquires power unit operation data, power grid transmission and distribution load data, and environmental monitoring data through a multi-dimensional sensor array to form a feature data set; a deep neural network module processes the feature data set based on a two-layer cross-validation mechanism to generate state evaluation data; a hierarchical reinforcement learning module constructs a multi-objective power dispatching model based on the state evaluation data and generates a control parameter matrix; an intelligent execution module receives the control signal and control parameter matrix output by the multi-objective power dispatching model; and the intelligent execution module performs parameter self-tuning according to the control parameter matrix through a built-in self-correction controller.

[0014] The beneficial effects of the present invention are as follows: the present invention proposes an integrated design scheme based on a multi-dimensional sensor array, integrates temperature sensing, voltage and current sampling, and meteorological sensing circuits on the same chip, realizes timing control through FPGA, and significantly improves the accuracy of data acquisition; designs an improved Transformer structure, deeply extracts power system features by integrating spatial attention network and temporal attention network, and adopts a three-stage strategy (self-supervised pre-training, supervised pre-training, end-to-end fine-tuning) to optimize model performance, thereby improving the accuracy of state assessment; proposes a multi-objective power dispatching model based on hierarchical reinforcement learning, and realizes dynamic balance adjustment of power supply stability and economy through collaborative optimization of the day-ahead dispatching sublayer and the real-time dispatching sublayer, combined with a fuzzy adaptive dynamic weight update mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 This is the system architecture diagram of the Internet-based artificial intelligence power dispatching management system.

[0017] Figure 2 This is a processing flow chart of the deep neural network module of the Internet-based artificial intelligence power dispatching and management system.

[0018] Figure 3 This is a flowchart of the hierarchical reinforcement learning module of the Internet-based artificial intelligence power dispatching management system. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0022] Example 1, reference Figure 1~Figure 3 , which is the first embodiment of the present invention, and this embodiment provides an artificial intelligence power dispatching management system based on the Internet. The system architecture diagram is as follows Figure 1 As shown, it includes a distributed data acquisition module, a deep neural network module, a hierarchical reinforcement learning module, an intelligent execution module and a security management module. Among them: The distributed data acquisition module obtains the operation data of the power unit, the transmission and distribution load data of the power grid, and the environmental monitoring data through a multi-dimensional sensor array to form a feature data set. The multi-dimensional sensor array adopts a modular separation design scheme, designs the temperature sensing circuit, the voltage and current sampling circuit, and the meteorological sensing circuit as independent functional modules, and realizes the synchronous acquisition of multiple parameters through FPGA. Specifically, the temperature sensing circuit adopts an RTD temperature sensor with a measurement range of -50℃~200℃, a three-wire wiring method and a standard current output, which is suitable for the temperature monitoring requirements of power equipment operation; the voltage and current sampling circuit uses a voltage transformer (PT) and a current transformer (CT) for isolated sampling to achieve a 0.5-level measurement accuracy, a voltage range of 0500V, a current range of 050A, and a sampling frequency that can be configured to 1Hz1kHz; the meteorological sensing circuit uses an industrial-grade digital sensor to monitor the ambient temperature (-20℃~70℃) and relative humidity (20%~90%RH).

[0023] The distributed data acquisition module is equipped with a data quality assessment unit, which uses a multi-level data screening algorithm to check the data integrity and validity of the feature data set, and detects outliers by setting multi-dimensional thresholds such as sampling frequency, signal amplitude and data timestamp. The data quality assessment adopts a three-level screening mechanism: the first level performs data integrity check, identifies data loss and transmission errors by detecting the continuity of the data packet header, CRC32 checksum and timestamp (timestamp accuracy is better than 1ms), and uses a linear interpolation algorithm based on historical data to compensate for the detected data loss; the second level performs range inspection, sets a reasonable threshold range based on the equipment nameplate parameters and historical operating data, and marks the over-limit data; the third level uses the sliding window method for mutation detection, with a window length of 60s, calculates the data change rate, and marks it when the change rate exceeds the preset threshold (usually 3 times the historical mean).

[0024] To improve data processing efficiency, this module configures an edge computing architecture to preprocess feature data sets, including data standardization, noise reduction, and compression. Data preprocessing uses a streaming processing architecture. The standardization process is based on the Z-score algorithm. A 24-hour sliding window is used to dynamically update normalization parameters, and the parameter update frequency is once an hour. Noise reduction uses a Kalman filter, and the process noise covariance and measurement noise covariance are configured according to the equipment operation characteristics. Data compression uses a piecewise linear approximation algorithm, and the compression ratio can reach 10:1. The maximum allowable error does not exceed 0.1% of the full scale, and the minimum segment length is 10 sampling points, while ensuring the retention of key feature points. To ensure the real-time performance of edge computing, a double buffering mechanism is used for data processing, and the processing delay is controlled within 100ms.

[0025] The processing flow chart of the deep neural network module is as follows Figure 2 As shown in the figure, the specific details are as follows: the feature data set is processed based on a two-layer cross-validation mechanism, which includes a time domain feature extraction layer and a frequency domain feature analysis layer. Specifically, the time domain feature extraction layer performs segmented processing on the feature data set, calculates statistics such as mean, standard deviation, and peak factor, and generates time domain statistical features; the frequency domain feature analysis layer performs wavelet transform on the time domain statistical features to obtain energy distribution characteristics in different frequency bands. The processing window length is set to 30 minutes according to the operating characteristics of the power equipment, and the overlap rate of adjacent windows is 25%. The statistical features include: basic statistics (mean, standard deviation, skewness, kurtosis), trend factors (first-order difference mean, variance), and pulse index (peak factor, margin factor, waveform factor) to characterize the dynamic characteristics of the data. The wavelet transform uses Morlet wavelet decomposition, and the frequency range covers 0.1Hz~1kHz, which can effectively capture the transient and steady-state characteristics in the power system.

[0026] The deep neural network module uses an improved Transformer structure to enhance the model's feature extraction capability by fusing the spatial attention network and the temporal attention network. The spatial attention network fuses the time domain statistical features and the frequency domain energy distribution features, and obtains the spatial feature vector by calculating the correlation weights between different monitoring points; the temporal attention network uses a gated loop structure to extract the time series pattern of the spatial feature vector and obtain key temporal features. The improved Transformer structure is optimized for the characteristics of the power system: the spatial attention network uses a 4-head attention mechanism, each with a dimension of 32, and a residual connection is added to enhance the stability of the model; the temporal attention network uses a bidirectional GRU structure with 128 hidden units, and Dropout (0.1) is added to prevent overfitting, and a gating mechanism is used to adaptively adjust the weights of long and short-term memory.

[0027] The steps of spatial attention network feature fusion are as follows: first extract time domain statistical features; then calculate frequency domain energy distribution features; then fuse features and calculate monitoring point correlation; finally generate spatial feature vectors through multi-head attention mechanism. The steps of temporal attention network processing are as follows: first process sequence information through bidirectional GRU; then use gating mechanism to calculate weights; finally weighted fusion to obtain key temporal features.

[0028] Subsequently, the Transformer structure inputs the spatial feature vector and key temporal features into the multi-head self-attention layer and the feedforward neural network to generate state evaluation data. The multi-head self-attention layer adopts an 8-head attention mechanism with a dimension of 64, and introduces position encoding to enhance the sequence modeling capability; the feedforward network adopts a two-layer structure [512, 256] and uses the GELU activation function.

[0029] In addition, the deep neural network module is equipped with a model evaluation unit, which uses a weighted cross entropy loss function to evaluate the accuracy of the state evaluation data online, and imposes higher penalty weights on important fault types. The model training adopts a three-stage strategy: first, the feature extraction network is pre-trained using a self-supervised learning method, and a contrast loss function is constructed to learn data representation; then, supervised pre-training is performed using typical fault cases in historical data; and finally, end-to-end fine-tuning training is performed. The training parameters are configured as follows: batch_size is 64, AdamW optimizer is used, weight decay coefficient is 0.01, learning rate adopts cosine annealing strategy, initial value is 0.001, minimum value is 0.00001, and the number of training rounds is 100. To prevent overfitting, an early stopping mechanism is used, and training is stopped if the performance of the validation set does not improve after 5 consecutive rounds.

[0030] The hierarchical reinforcement learning module builds a multi-objective power dispatch model based on the state assessment data. The hierarchical reinforcement learning module flow chart is as follows: Figure 3As shown in the figure, the module first extracts the power load characteristics, power supply stability characteristics and economic cost characteristics in the state assessment data in a hierarchical manner based on the sampling strategy, and constructs the feature vector space. The sampling strategy adopts the hierarchical importance sampling method: the load characteristics are stratified according to the power consumption type (industrial, commercial, and civil), with a sampling weight ratio of 4:3:3, and are secondary stratified according to the time distribution characteristics (working days / holidays); the power supply stability characteristics focus on sampling key state points such as voltage over-limit (±7% rated value) and power fluctuation (rate of change>10% / min), and the sampling frequency is adaptively adjusted with the fault level; the economic cost characteristics are stratified based on the peak and valley electricity price periods (1.5 times the benchmark electricity price during the peak period and 0.7 times the benchmark electricity price during the valley period). The feature vector is stored in a sparse coding manner, with a vector dimension of 256, and L1 regularization is used to control the sparsity within 20%.

[0031] The hierarchical reinforcement learning module also sets up a hierarchical scheduling engine, including a day-ahead scheduling sublayer and a real-time scheduling sublayer. The day-ahead scheduling sublayer is based on a dual DQN network (DDQN), which maps the power supply stability characteristics and economic cost characteristics into power supply stability constraint functions (considering indicators such as voltage qualification rate and power factor) and operation economic constraint functions (including power generation cost, network loss cost and regulation cost), respectively; the real-time scheduling sublayer performs short-term load forecasting based on the LSTM model (forecasting period of 4 hours, step length of 15 minutes), and uses the A2C (Advantage Actor-Critic) method with a baseline to build a dual-objective optimization model. The DDQN network adopts a dual network structure, including 3 hidden layers (dimensions are 128, 256, and 128, respectively), and uses the LeakyReLU activation function ( ), the target network is soft updated every 100 steps ( In the A2C architecture, the Actor network outputs the probability distribution of the scheduling action (using the Softmax layer), and the Critic network evaluates the state-action value. Both networks use a 3-layer fully connected structure, and the middle layer uses Batch Normalization to improve training stability.

[0032] The hierarchical reinforcement learning module is also equipped with a dynamic weight update circuit. According to the real-time load state and prediction deviation, the weight parameters of power supply stability and economy in the dual-objective optimization model are dynamically adjusted based on the adaptive algorithm, and the neural network parameters are updated based on the gradient descent algorithm and the back propagation algorithm to obtain the control parameter matrix. The weight update adopts the fuzzy adaptive method: 7 fuzzy language variables are designed to describe the load state (extremely low [<30%], relatively low [30-50%], slightly low [50-70%], normal [70-90%], slightly high [90-110%], relatively high [110-130%], extremely high [>130%]), 5 language variables are designed to describe the prediction deviation (very small [<2%], small [2-5%], medium [5-10%], large [10-15%], very large [>15%]), and 35 adaptive fuzzy rules are constructed based on expert experience and historical data analysis. The adjustment range of the weight parameter is [0.3, 0.7], and the sliding average method is used to smooth the weight changes. The time window is 30 minutes.

[0033] In addition, the control parameter matrix includes the control gain coefficient, the integral time constant and the differential time constant. The hierarchical reinforcement learning module is equipped with a dispatch strategy optimization unit, which continuously optimizes the multi-objective power dispatch model based on the experience replay mechanism. The experience replay adopts a priority sampling strategy, the replay pool capacity is 10,000, the priority is calculated based on time decay and the absolute value of the reward, and the batch size of each sampling is 32.

[0034] The intelligent execution module receives the control signal and control parameter matrix output by the multi-objective power dispatching model. The module sets a control signal processing unit to parse and classify the received control signal; sets a parameter matrix analysis unit to perform feature decomposition on the control parameter matrix and obtain the control parameters. Control signals are divided into three categories according to execution priority: emergency control signals (need to respond within 50ms), ordinary control signals (response within 1s) and planned control signals (response within 10s). Among them, the emergency control signal adopts an interrupt trigger mechanism and is directly output through the hardware acceleration channel, and the response delay does not exceed 20ms. The singular value decomposition (SVD) method is used for feature decomposition, and the eigenvalues ​​with a cumulative contribution rate of 99.5% are retained. The adaptive threshold method is used for eigenvalue screening, and the contribution rate threshold is dynamically adjusted according to the system operation status, and it is reduced to 95% during power grid disturbance to improve the calculation speed.

[0035] The intelligent execution module also sets a parameter decomposer to decompose the control parameter matrix into a control target parameter group and a constraint parameter group, and normalizes these parameter groups. Parameter decomposition is based on physical meaning. The control target parameter group contains the target values ​​of operating indicators such as voltage, power, and frequency, and the constraint parameter group contains constraints such as equipment operating limits and safety margins. A cross-validation mechanism is set between parameter groups to identify potential correlations between parameters through pattern recognition algorithms to avoid constraint conflicts. Normalization uses piecewise linear transformation to ensure that parameter values ​​fall within the effective range of the controller. The normalization process introduces a dynamic dead zone to suppress small fluctuations and improve system stability.

[0036] The intelligent execution module has a built-in self-correcting controller, and a control parameter mapping rule base is established based on the fuzzy reasoning mechanism, including the control error magnitude and error change rate. The rule base has an online learning function, which can automatically adjust the rule weight according to the control effect and eliminate inefficient rules. The self-correcting controller calculates the control parameter correction amount according to the control error magnitude and error change rate, which is used to dynamically adjust the control gain coefficient, integral time constant and differential time constant. The fuzzy rule base contains 49 rules (7×7 rule matrix), the input variable error magnitude and change rate are both based on 7 fuzzy sets, the output variable correction amount is based on 9 fuzzy sets, and the defuzzification adopts the centroid method. The membership function of the fuzzy set adopts adaptive deformation technology, and the function shape can be dynamically adjusted according to the control accuracy requirements.

[0037] The self-correcting controller sets the parameter tuning module, based on the adaptive PID control algorithm, dynamically adjusts the control parameters according to the control parameter correction amount to complete the parameter self-tuning. The adjustment range of PID parameters is: K p ∈[0.3,2.5],K i ∈[0.05,1.2], K d ∈[0.02,0.8]. The forgetting factor mechanism is introduced into the parameter adjustment process to reduce the impact of historical data and improve the system's adaptability to new features.

[0038] The self-correcting controller is equipped with a feedback compensation circuit to correct the control deviation in real time. The module is also equipped with an execution effect evaluation unit, which uses a sliding time window algorithm to monitor and evaluate the execution effect of the control instruction in real time. The feedback compensation adopts a feedforward-feedback structure. The feedforward channel predicts the interference effect based on the system model, and the feedback channel uses a first-order lag link for filtering with an adjustable time constant. The feedforward channel integrates a neural network predictor, which can learn the dynamic characteristics of the system online and improve the interference suppression effect. The evaluation window length is 3 minutes, the overlap rate is 75%, and the control index arrival and stability index are calculated. The evaluation index adopts a multi-objective dynamic weighting method to automatically adjust the weight of each index according to the importance of the control task.

[0039] In addition, the self-tuning controller performs parameter self-tuning according to the control parameter matrix, including: the self-tuning controller establishes a parameter mapping relationship based on the fuzzy reasoning mechanism, and the parameter mapping relationship includes a control gain factor and a time constant factor; the mapping relationship is established by using the recursive least squares method to achieve rapid convergence of parameter estimation; the self-tuning controller optimizes and adjusts the control signal according to the parameter mapping relationship; the optimization process introduces a particle swarm algorithm to solve the optimal control quantity under the premise of ensuring real-time performance; the self-tuning controller uses an incremental tuning method to adaptively adjust the control gain factor and the time constant factor; an anti-saturation integrator is set in the adjustment process to avoid system oscillation caused by integral saturation; the self-tuning controller generates a corrected control instruction based on the incremental tuning method. The control instruction generation adopts a model predictive control strategy to predictively consider the multi-step control effect.

[0040] The security management module conducts security monitoring of the entire system operation process, including data transmission encryption, access rights management, fault diagnosis and emergency handling functions. Data transmission adopts the AES-256 encryption algorithm, and the key is exchanged through the elliptic curve cryptography algorithm (ECC); the key update adopts a dynamic time window mechanism, and the window length is adaptively adjusted according to the system load to ensure a balance between security and performance. The access rights adopt the role-based access control (RBAC) model and set 4 levels of authority; the authority management introduces a behavioral analysis engine, which identifies abnormal operation modes through deep learning algorithms to achieve active defense. Fault diagnosis combines expert systems and deep learning methods, and the diagnostic accuracy is not less than 95%. The diagnostic model adopts an integrated learning framework, integrates the prediction results of decision trees, support vector machines and neural networks, and introduces a confidence assessment mechanism.

[0041] The security management module adopts a distributed architecture, deploys security protection units at each key node of the system, and realizes multi-level and all-round security protection. A P2P communication network is established between each protection unit, and blockchain technology is used to record security events to ensure reliable traceability. The security protection unit adopts a four-layer protection architecture: the first layer is physical layer protection, which implements physical isolation and equipment redundancy; physical protection adopts intelligent sensor network, and realizes early warning of intrusion threats through multi-source information fusion algorithm. The second layer is network layer protection, deploying firewalls and intrusion detection systems; the intrusion detection system integrates abnormal traffic analysis and deep packet inspection technology to realize heuristic protection against zero-day vulnerabilities. The third layer is application layer protection, which realizes identity authentication and operation audit; identity authentication combines biometric recognition and dynamic token technology, and adopts adaptive authentication strategy. The fourth layer is data layer protection, which performs data encryption and integrity verification. Data integrity verification uses distributed hash tables to store characteristic values, and designs differential backup mechanism to realize rapid data recovery.

[0042] The security management module monitors the system operation status in real time, promptly discovers and handles potential safety hazards, and ensures the stable operation of the system. The monitoring system adopts a hierarchical adaptive sampling strategy to dynamically adjust the sampling frequency for parameters of different importance levels. Real-time monitoring is based on the state machine model, setting 24 system states and 48 state transition conditions. Any abnormal state transition will trigger the alarm mechanism. The state machine model integrates a fuzzy temporal reasoning engine to handle the uncertainty in the state transition process and improve the accuracy of the alarm. The alarm mechanism adopts a multi-level early warning strategy, automatically adjusts the response level according to the threat level, and has an intelligent diversion function to avoid alarm storms. The system introduces a self-healing mechanism, optimizes the fault recovery strategy through a reinforcement learning algorithm, and realizes automatic repair of the system.

[0043] Furthermore, the present embodiment also provides an artificial intelligence power dispatching management method based on the Internet, including a distributed data acquisition module that acquires power unit operation data, power grid transmission and distribution load data, and environmental monitoring data through a multi-dimensional sensor array to form a feature data set; a deep neural network module processes the feature data set based on a two-layer cross-validation mechanism to generate state evaluation data; a hierarchical reinforcement learning module constructs a multi-objective power dispatching model based on the state evaluation data and generates a control parameter matrix; an intelligent execution module receives a control signal and a control parameter matrix output by the multi-objective power dispatching model; and the intelligent execution module performs parameter self-tuning according to the control parameter matrix through a built-in self-correcting controller. Example 2, reference Figure 1~Figure 3 , which is the second embodiment of the present invention, provides an artificial intelligence power dispatching and management system based on the Internet. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0044] To verify the performance of the artificial intelligence power dispatching and management system, the research team conducted a three-month simulation experiment at a 220kV substation of a provincial power grid company. The experimental platform uses an Intel Xeon Gold 5320 processor, 128GB memory, and is equipped with an NVIDIA A5000 GPU for deep learning computing acceleration.

[0045] The experiment selected two 110kV substations and eight distribution stations under the substation as test objects, and deployed a total of 324 measurement points, including key parameters such as voltage, current, active power, and reactive power. The sampling frequency of the distributed data acquisition module was set to 50Hz, the data preprocessing window length was 60s, and the compression ratio was configured to 6:1. The deep neural network module used a 30-minute processing window, the batch size was set to 32, the number of training rounds was 80, and the validation set ratio was 20%.

[0046] In terms of load prediction, experimental data show that the average prediction error of the system is 3.56%, which is significantly improved compared with the 5.12% of the traditional method. Especially in the case of sudden load changes, the system can complete the model parameter adjustment within an average of 2.3 seconds and control the prediction error within 4.8%. In terms of power supply reliability indicators, the system has increased the voltage qualification rate from the original 94.8% to 97.2%, and the power factor from 0.91 to 0.94.

[0047] The system also shows certain advantages in terms of economic benefits. By optimizing the dispatching strategy, the system reduced the network loss rate from 2.3% to 1.95%, and the monthly operating cost was reduced by about 5.8%. In terms of dispatching response time, the average response delay of emergency control signals is 45ms, ordinary control signals is 0.92s, and planned control signals is 8.5s, which basically meets the design index requirements.

[0048] The performance comparison between the present invention and the prior art is shown in Table 1.

[0049] Table 1 Comparison of main performance indicators of the solution of the present invention and the prior art Evaluation indicators Traditional methods The present invention Average load forecast error 5.12% 3.56% Load mutation response time 4.8s 2.3s Voltage pass rate 94.8% 97.2% Network loss rate 2.3% 1.95% Operating cost reduction ratio - 5.8% Emergency control signal response delay 95ms 45ms Fault diagnosis accuracy 91% 94.5% System stability 1.8% 0.9% From the comparative data in Table 1, it can be seen that the method of the present invention is superior to the existing technology in all key performance indicators. In particular, the performance is significantly improved in terms of load prediction accuracy, system response speed and operating efficiency. For example, the average error of load prediction is reduced by 30.5% compared with the traditional method, the response delay of emergency control signal is shortened by 52.6%, and the operating cost is reduced by 5.8%. These indicators show that the present invention has good application value.

[0050] In the system security test, the penetration test method was used to conduct a two-week security assessment of the system. The results showed that the system can effectively defend against 97.5% of known attack types, and the average response time of the security protection unit is 125ms. In terms of data transmission, the average delay of the AES-256 encryption algorithm is 3.8ms, and the accuracy of data integrity verification reaches 99.95%. After 3 months of continuous operation testing, the annual failure rate of the system has dropped to 0.9%, which is better than the industry average.

[0051] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An artificial intelligence power dispatching and management system based on the Internet, characterized by: It includes distributed data acquisition module, deep neural network module, hierarchical reinforcement learning module, intelligent execution module and security management module, among which: The distributed data acquisition module acquires the power unit operation data, power grid transmission and distribution load data and environmental monitoring data through a multi-dimensional sensor array to form a feature data set; the multi-dimensional sensor array integrates a temperature sensing circuit, a voltage and current sampling circuit and a meteorological sensing circuit; The deep neural network module processes the feature data set based on a two-layer cross-validation mechanism to generate state evaluation data; the two-layer cross-validation mechanism includes a time domain feature extraction layer and a frequency domain feature analysis layer; the deep neural network module applies a Transformer structure, and the Transformer structure integrates a spatial attention network and a temporal attention network; The hierarchical reinforcement learning module constructs a multi-objective power dispatching model based on the state evaluation data; the hierarchical reinforcement learning module configures a dynamic weight update circuit and generates a control parameter matrix based on an online learning algorithm; The intelligent execution module receives the control signal output by the multi-objective power dispatching model and the control parameter matrix, and the intelligent execution module has a built-in self-correcting controller, and the self-correcting controller performs parameter self-tuning according to the control parameter matrix; The security management module performs security monitoring on the entire process of system operation.

2. The Internet-based artificial intelligence power dispatching and management system according to claim 1, characterized in that: The distributed data acquisition module configures an edge computing architecture to pre-process the feature data set; The two-layer cross-validation mechanism sets a data screening threshold for filtering abnormal data in the feature data set; The self-correcting controller is provided with a feedback compensation circuit to correct the control deviation in real time.

3. The Internet-based artificial intelligence power dispatching and management system according to claim 1, characterized in that: The deep neural network module includes the following steps: The time domain feature extraction layer performs segmentation processing on the feature data set to generate time domain statistical features; The frequency domain feature analysis layer performs wavelet transform on the time domain statistical features to obtain frequency domain energy distribution features; The spatial attention network fuses the time domain statistical features and the frequency domain energy distribution features, calculates the correlation weights of different monitoring points, and obtains a spatial feature vector; The temporal attention network extracts the time series pattern of the spatial feature vector to obtain key temporal features; The Transformer structure inputs the spatial feature vector and the key temporal features into a multi-head self-attention layer and a feedforward neural network to generate state evaluation data.

4. The Internet-based artificial intelligence power dispatching and management system according to claim 1, characterized in that: The hierarchical reinforcement learning module includes: Based on the sampling strategy, the power load characteristics, power supply stability characteristics and economic cost characteristics in the state assessment data are extracted in layers to construct a feature vector space; Setting a hierarchical scheduling engine, wherein the hierarchical scheduling engine includes a day-ahead scheduling sublayer and a real-time scheduling sublayer; The day-ahead scheduling sublayer adopts a dual DQN network to map the power supply stability characteristics and economic cost characteristics into a power supply stability constraint function and an operation economy constraint function respectively; The real-time scheduling sublayer performs load prediction based on the LSTM model and builds a dual-objective optimization model; The dynamic weight updating circuit adjusts the weight parameters of power supply stability and economy in the dual-objective optimization model according to the real-time load state and the prediction deviation, and updates the model parameters based on the gradient descent algorithm and the back propagation algorithm to obtain the control parameter matrix; The control parameter matrix includes a control gain coefficient, an integral time constant and a differential time constant.

5. The Internet-based artificial intelligence power dispatching and management system according to claim 1, characterized in that: The self-tuning controller performs parameter self-tuning according to the control parameter matrix, including: The self-correcting controller establishes a parameter mapping relationship based on a fuzzy inference mechanism, and the parameter mapping relationship includes a control gain factor and a time constant factor; The self-correcting controller optimizes and adjusts the control signal according to the parameter mapping relationship; The self-correcting controller uses an incremental tuning method to adaptively adjust the control gain factor and the time constant factor; The self-correcting controller generates corrected control instructions based on the incremental tuning method.

6. The Internet-based artificial intelligence power dispatching and management system according to claim 1, characterized in that: The intelligent execution module includes a control signal processing unit and a parameter matrix analysis unit, wherein: The control signal processing unit receives the control signal output by the multi-objective power dispatching model, and performs signal analysis and classified storage; The parameter matrix analysis unit performs characteristic decomposition on the control parameter matrix to obtain a control gain coefficient, an integral time constant and a differential time constant; The self-correcting controller establishes a control parameter mapping rule base based on a fuzzy inference mechanism, and the control parameter mapping rule base includes a control error magnitude and an error change rate; The self-correcting controller calculates a control parameter correction amount according to the control error magnitude and the error change rate, and the control parameter correction amount is used to dynamically adjust the control gain coefficient, the integral time constant and the differential time constant; The self-correcting controller is provided with a parameter setting module, and the parameter setting module dynamically adjusts the control parameter according to the control parameter correction amount to complete parameter self-tuning.

7. The Internet-based artificial intelligence power dispatching and management system according to claim 1, characterized in that: The distributed data acquisition module is provided with a data quality assessment unit, and the data quality assessment unit adopts a multi-level data screening algorithm to perform data integrity check and validity verification on the feature data set; The deep neural network module is provided with a model evaluation unit, which performs online evaluation on the accuracy of the state evaluation data based on a cross entropy loss function.

8. An Internet-based artificial intelligence power dispatching management method, based on the Internet-based artificial intelligence power dispatching management system according to any one of claims 1 to 7, characterized in that: Also includes, The distributed data acquisition module acquires power unit operation data, power grid transmission and distribution load data and environmental monitoring data through a multi-dimensional sensor array to form a feature data set; The deep neural network module processes the feature data set based on a two-layer cross-validation mechanism to generate state assessment data; The hierarchical reinforcement learning module constructs a multi-objective power dispatch model based on the state evaluation data and generates a control parameter matrix; The intelligent execution module receives the control signal output by the multi-objective power dispatching model and the control parameter matrix; The intelligent execution module performs parameter self-tuning according to the control parameter matrix through a built-in self-correcting controller.

Images