An AI-based intelligent energy gateway intelligent scheduling and optimization control system

Abstract

The application discloses an AI-based intelligent energy gateway intelligent scheduling and optimization control system, which comprises an energy access sheet, a data processing sheet, an AI decision sheet and a control execution sheet. The energy access sheet collects and uniformly converts equipment operation data of different protocols, and the data processing sheet performs differential processing and abnormality inspection; the AI decision sheet adopts a time sequence convolution network (TCN) to perform load prediction and uses a deep reinforcement learning network (DQN) to calculate an optimal power distribution scheme; and the control execution sheet is responsible for executing control instructions and monitoring states. The application solves technical problems of a traditional energy gateway in aspects of multi-protocol equipment access, data processing and intelligent decision optimization, predicts load changes through adaptive learning, optimizes power distribution based on dynamic coupling relationships among equipment, and improves energy utilization efficiency and system operation stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart energy management, and particularly relates to an AI-based smart energy gateway intelligent scheduling and optimization control system. BACKGROUND

[0002] With the large-scale access of distributed energy devices such as photovoltaic and energy storage, the smart energy gateway needs to realize the collaborative scheduling and optimization control of power generation devices, energy storage devices and power consumption loads. In particular, in large energy consumption scenarios such as industrial parks and commercial buildings, load prediction accuracy and device collaboration efficiency directly affect the economic benefits of the system. The energy gateway products on the market currently have the following technical problems in actual application:

[0003] 1. Low load prediction accuracy: existing systems mostly use simple time series analysis or statistical regression methods to predict load changes, which have obvious shortcomings: on the one hand, they cannot effectively capture the long-term dependence characteristics of load data, resulting in cumulative errors in medium and long-term prediction results; on the other hand, they lack the ability to identify complex patterns such as load mutations and seasonal fluctuations, affecting the accuracy of the prediction. At the same time, the prediction model lacks the ability to deeply mine historical data and is difficult to adapt to the dynamic changes of load patterns.

[0004] 2. Sub-optimal device collaboration decision: traditional scheduling methods mainly allocate power based on simple fixed rules such as "peak load shifting" or "peak valley arbitrage", which have great limitations. Specifically, first, the dynamic matching relationship between power generation devices and energy storage devices is not fully considered, resulting in inefficient use of excess power; second, the capacity limit and charge-discharge characteristics of energy storage devices are not fully considered, which can easily lead to overuse or idling of energy storage resources; third, there is no response characteristic model of load changes and energy storage adjustment, making it difficult to achieve precise adjustment of load fluctuations. These problems lead to low overall system operation efficiency and cannot achieve optimal collaboration of energy devices.

[0005] These technical problems seriously restrict the effectiveness of smart energy gateways in actual application, and there is an urgent need to develop a new energy gateway system with high-precision load prediction and intelligent collaborative decision-making capabilities. SUMMARY

[0006] The problem to be solved by the present application is to provide a smart energy gateway intelligent scheduling and optimization control system with high-precision load prediction and intelligent collaborative decision-making capabilities.

[0007] To solve the above technical problems, the technical solution adopted by the present application is: an AI-based smart energy gateway intelligent scheduling and optimization control system, characterized in that it comprises:

[0008] Energy access piece: used to access and collect real-time operation data of power generation equipment, energy storage devices and power consumption loads supporting different communication protocols, and uniformly convert data of different protocols into system standard data format containing device identification, timestamp, operation data and state information;

[0009] Data processing piece: used to perform differential processing on operation data in standard data format by calculating the change amount of adjacent sampling data, perform anomaly detection on data exceeding the preset threshold, and distribute valid data to the AI decision piece;

[0010] AI decision piece: uses time series convolution network TCN to build a load prediction model, and uses deep reinforcement learning network DQN to calculate the optimal power allocation scheme based on the prediction result;

[0011] Control execution piece: used to convert the power allocation scheme into device control instructions and execute them while monitoring the instruction execution status.

[0012] Further technical solutions are that the energy access piece includes a gateway basic module, a protocol analysis module, a protocol conversion module and a data cache module:

[0013] The gateway basic module is used to establish and maintain a bidirectional communication connection with power generation equipment, energy storage devices and power consumption equipment, support multiple device communication protocols, and realize data acquisition and control instruction transmission;

[0014] The protocol analysis module is used to analyze device communication frames of different protocols and extract voltage, current and power operation data;

[0015] The protocol conversion module is used to uniformly convert the analyzed device data into a system standard data format, including a device identification field, a timestamp field, an operation data field and a state field, wherein the device identification field includes device type and device communication address, the timestamp field is a UNIX timestamp of the data acquisition time, the operation data field includes device voltage value, current value and power value, and the state field includes device operation state and communication state;

[0016] Further technical solutions are that the data processing piece includes a data difference unit, a data verification unit and a data distribution unit:

[0017] The data difference unit is used to perform differential processing on operation data in the standard data format cached by the data cache module, set differential thresholds of voltage, current and power according to device type, calculate the change amount of adjacent sampling data and compare it with the corresponding preset threshold, and when the change amount exceeds the preset threshold, transmit the latest data to the data verification unit for processing;

[0018] A data inspection unit is configured to inspect the data transmitted by the data difference unit, determine whether the voltage, current and power data are within the rated value of the device, and determine the data as valid data when the voltage, current and power data are within the rated value of the device and the data transmission is normal.

[0019] A data distribution unit is configured to transmit the qualified data in real time to the AI decision piece.

[0020] Further, the technical scheme is that the time sequence convolution network TCN is configured to build a load prediction model, which specifically includes:

[0021] A historical operation database is established to store the historical voltage, current and power data of the power generation device, energy storage device and power load.

[0022] Based on the historical operation data, a time sequence convolution network TCN containing multiple layers of cavity convolution layers is built to analyze the long-term dependence of the historical data through an incremental cavity rate and a residual connection structure.

[0023] The predicted load value, load change trend feature and load fluctuation range are output.

[0024] Further, the deep reinforcement learning network DQN includes an input layer, a hidden layer and an output layer.

[0025] The input layer is configured to build an input vector containing the voltage, current and power operation data of the power generation device, energy storage device and power load transmitted by the data distribution unit, and the predicted load value, load change trend feature and load fluctuation range output by the time sequence convolution network TCN.

[0026] The hidden layer is configured to build a state vector from the real-time operation data and the prediction result, process the state vector through a neural network weight and an activation function, automatically extract a dynamic coupling relationship feature between devices, and the coupling relationship feature includes a matching feature of the power generation device and the charging demand of the energy storage device, which is used to represent the excess power that can be used for energy storage; a constraint feature of the state of charge of the energy storage device and the discharge power, which is used to represent the available capacity and maximum charge and discharge power of the energy storage device; and a response feature of the power load change and the charging and discharging adjustment of the energy storage device, which is used to represent the demand amount of the load fluctuation to the energy storage adjustment.

[0027] The output layer is configured to use the dynamic coupling relationship feature to guide the power distribution decision, calculate the optimal power distribution scheme based on the deep reinforcement learning network DQN, perform a nonlinear transformation calculation value evaluation value on the state vector and the dynamic coupling relationship feature by using a neural network, and perform online learning on the value evaluation network by using a time sequence difference algorithm, and the formula of the time sequence difference algorithm is:

[0028] Q(st h t a t )=r t +γQ(s t+1 h t+1 a t+1 )

[0029] Wherein, s t is a state vector, h t is a dynamic coupling relationship feature, a t is a power allocation scheme, r t is a system reward value of the current power allocation scheme, gamma is a discount factor for balancing current decision and future impact, Q(s t h t a t ) is a value evaluation value of the power allocation scheme under the current state, and Q(s t+1 h t+1 a t+1 ) is a value evaluation of the next state.

[0030] Based on the value evaluation result, an epsilon-greedy strategy is used to select the power allocation scheme with the highest value, and the power allocation values of the power generation equipment, the charge and discharge power allocation values of the energy storage device and the power distribution power allocation values of the power consumption load are output.

[0031] Further technical solutions are as follows: the control execution piece includes an instruction conversion unit and a control execution unit.

[0032] The instruction conversion unit is used to convert the power values output by the AI decision piece into communication control instructions of various devices: for the power generation equipment, the power value is converted into a power generation power instruction conforming to its communication protocol; for the energy storage device, the positive power is converted into a discharge power instruction, and the negative power is converted into a charging power instruction; for the power consumption load, the power value is converted into a power consumption power limit instruction.

[0033] The control execution unit is used to encapsulate the converted control instructions into messages according to the requirements of the communication protocols of various devices and send them to the corresponding devices through the gateway basic module, confirm the execution state of the control instructions based on the collected device power, voltage and current operating state feedback information, and resend the instructions when an instruction execution exception is detected.

[0034] The beneficial effects of the present application are as follows:

[0035] (1) The application adopts a time sequence convolution network TCN for load prediction, significantly improves the capture ability of long-term dependence characteristics of load data through a multi-layer hollow convolution structure, and effectively alleviates the gradient disappearance problem of a deep network by using residual connection. Compared with traditional prediction methods, the application can more accurately identify the seasonal fluctuation and mutation characteristics of the load, and the prediction accuracy is significantly improved;

[0036] (2) The application innovatively constructs a power optimization distribution scheme based on a deep reinforcement learning network DQN, extracts dynamic coupling relationships such as power generation-energy storage matching characteristics, energy storage capacity constraint characteristics and load-energy storage response characteristics, and realizes accurate modeling of the system state. Online learning is performed by using a time sequence difference algorithm, and the system can adaptively adjust the decision-making strategy according to the real-time state, overcoming the poor adaptability of traditional fixed rule scheduling methods to environmental changes; BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 It is an AI-based intelligent energy gateway intelligent scheduling and optimization control system structure diagram.

[0038] Figure 2 It is a TCN load prediction model network structure diagram.

[0039] Figure 3 It is a DQN deep reinforcement learning network structure diagram.

[0040] Figure 4 It is a control instruction execution flowchart. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely in conjunction with the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0042] The technical scheme of the application will be described in detail in the following specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described again in some embodiments.

[0043] As Figure 1As shown, the AI-based intelligent energy gateway scheduling and optimization control system provided by the present application comprises an energy access piece, a data processing piece, an AI decision piece and a control execution piece. The energy access piece is used to solve the access problem of different communication protocol devices, realize unified collection and standardized conversion of multi-protocol device data; the data processing piece is responsible for differential processing and validity verification of collected data; the AI decision piece realizes load prediction and optimal power distribution based on deep learning; and the control execution piece is responsible for converting the decision result into specific device control instructions and monitoring the execution.

[0044] The energy access piece comprises a gateway basic module, a protocol analysis module, a protocol conversion module and a data cache module. The gateway basic module is responsible for establishing and maintaining a bidirectional communication connection with power generation equipment, energy storage devices and power load equipment. The module supports multiple communication protocols including Modbus RTU / TCP protocol, IEC 60870-5-101 / 104 protocol and DNP3 protocol, can automatically identify the communication protocol type of the device, and establish a corresponding communication connection according to the preset communication parameter template.

[0045] The protocol analysis module receives the original communication data collected by the gateway basic module, analyzes according to the frame format of different protocols, and extracts device operation data including voltage, current and power. In order to realize unified management of system data, the protocol conversion module converts the analyzed device data into a standard data format containing a device identification field, a timestamp field, an operation data field and a state field. The device identification field contains the device type and the communication address, the timestamp field adopts the UNIX timestamp format, the operation data field contains the voltage value, the current value and the power value, and the state field contains the device operation state and the communication state. The data cache module adopts a ring buffer structure, locally caches the converted standard format data, and realizes the data persistence function.

[0046] The data processing piece comprises a data difference unit, a data verification unit and a data distribution unit. In the specific implementation process, these units work cooperatively to ensure that the system obtains high-quality effective data.

[0047] In the specific implementation of the data difference unit, the system sets the corresponding difference threshold according to different device types. For power generation equipment, the voltage difference threshold can be set to ±2% of the rated voltage, for example, when the rated voltage is 380V, the difference threshold is ±7.6V; the current difference threshold is set to ±5% of the rated current; and the power difference threshold is set to ±3% of the rated power. For energy storage devices, due to the dynamic characteristics of the charging and discharging process, the power difference threshold can be appropriately relaxed to ±5% of the rated power. For power load, considering the randomness of load fluctuation, the power difference threshold can be set to ±4% of the rated power.

[0048] The specific implementation process of the differential processing is: first, the standard format data in the data cache module is obtained, for each newly collected data point, the change between it and the last sampling period data is calculated. Taking power data as an example, if the current sampling value is P1, the last period sampling value is P0, then the change is calculated as:

[0049] ΔP = |P1 - P0|

[0050] The calculated change ΔP is compared with the preset threshold value corresponding to the device type, when the change exceeds the preset threshold value, the latest data is transmitted to the data verification unit for processing.

[0051] After receiving the data transmitted by the differential unit, the data verification unit first performs the rated value range check. In specific implementation, the voltage effective range is set to 0.9U n to 1.1U n , wherein Un is the rated voltage of the device; the current effective range is 0 to 1.2I n , wherein I n is the rated current of the device; the power effective range is 0 to P n , wherein P n is the rated power of the device. In addition to the numerical range check, it is also necessary to verify whether the timestamp of the data is continuous, whether the communication state is normal, whether the device state is normal, etc. Only the data that meets all the verification conditions will be judged as valid data.

[0052] For data consistency verification, the system verifies by calculating the deviation of the product of voltage and current from the power value, and judges as abnormal data when the deviation exceeds 5% of the rated power. Data persistence adopts relational database storage, establishes a data table structure containing timestamp, device identification, voltage, current, power and state, and ensures data uniqueness through the joint primary key of timestamp and device identification. At the same time, the system establishes a composite index of storage data time, device type and data type to improve data query efficiency.

[0053] In a specific embodiment, if the rated parameters of a power generation device are: rated voltage 380V, rated current 100A, rated power 50kW, then the effective range of data verification is: voltage 342V-418V, current 0-120A, power 0-50kW. When the collected data is: voltage 375V, current 80A, power 30kW, since each parameter is within the effective range, and the device communication is normal, the data set will be judged as valid data.

[0054] The data distribution unit employs a publish-subscribe model for real-time data transmission. In practice, the unit establishes an independent data queue for each data type, with a queue length of up to 100, and updates data using a first-in, first-out (FIFO) approach. When new, valid data arrives, the system adds it to the corresponding data queue and simultaneously notifies the AI ​​decision-making unit that new data is available. To ensure reliable data transmission, the system uses an acknowledgment mechanism; the AI ​​decision-making unit must return an acknowledgment message after successfully receiving data. If no acknowledgment is received within a preset timeout period (e.g., 100ms), the data distribution unit will automatically resend the data.

[0055] In one specific embodiment of the present invention, the AI ​​decision chip uses the TCN load prediction model and the DQN deep reinforcement learning network to achieve intelligent scheduling and optimization control.

[0056] The TCN load forecasting model first establishes a historical operating database. In practice, the storage period for historical operating data can be set to 3 months, with a sampling frequency of once every 15 minutes. Voltage, current, and power data of power generation equipment, energy storage devices, and electrical loads are recorded and stored. This sampling frequency setting effectively balances the needs for data accuracy and storage space. Before storage, the data undergoes standardization, normalizing various data types to the [0,1] interval to improve the stability of subsequent model training.

[0057] like Figure 2 As shown, the TCN load prediction model network in this embodiment adopts a three-layer dilated convolutional structure with dilation rates set to 1, 2, and 4 respectively. This increasing dilation rate setting allows the model to expand its receptive field layer by layer, effectively capturing load change characteristics at different time scales. Each dilated convolutional layer is connected by residual connections, specifically by directly adding the input and output of each layer. This residual connection implementation facilitates the training process of deep networks. Regarding the TCN network structure, the specific parameter configurations for each dilated convolutional layer are as follows: the first dilated convolutional layer uses 64 output channels, the second layer uses 128 output channels, and the third layer uses 256 output channels, with a kernel size of 3 for all layers. To ensure causality, all convolutional layers use causal padding. During model training, a combination of MSE loss and Huber loss is used as the loss function, with a weight ratio of 7:3, to ensure robustness to both normal and abnormal samples. The optimizer uses the Adam algorithm with an initial learning rate of 0.001 and employs a learning rate decay strategy, halving the learning rate if the validation loss shows no improvement within 5 epochs. An early stopping mechanism is implemented during training, halting training if the validation loss shows no improvement within 15 epochs.

[0058] For long-term dependency analysis, the input sequence is set to 24 hours of historical data, containing about 96 data points in implementation. The model output includes load prediction values for the next 24 hours, as well as load change trend characteristics (up, stable, or down) and upper and lower limit values of load fluctuation range.

[0059] As shown in Figure 3 , in the implementation of the DQN deep reinforcement learning network, in terms of DQN network structure, a four-layer fully connected structure is adopted: the input layer dimension is 39 (including 12-dimensional running data, 24-dimensional prediction data, and 3-dimensional coupling features), the number of neurons in the middle two hidden layers is 256 and 512 respectively, and the output layer corresponds to the action space dimension 2100 (considering the combination of 10 grades of power generation, 21 grades of energy storage, and 10 grades of load power). To improve training stability, the network introduces an experience replay mechanism, with a replay pool capacity of 100000 and a batch size of 32 for each sampling. The target network is updated every 200 steps, using a soft update method with an update coefficient τ = 0.001.

[0060] The input layer receives two types of data: one is the real-time operating data transmitted by the data distribution unit, including the real-time power generation of the power generation equipment, the current state of charge and charging / discharging power of the energy storage device, and the real-time power consumption of the power load; the second is the prediction results output by the TCN model. These data are also standardized before input.

[0061] The hidden layer automatically extracts three types of dynamic coupling relationship features through neural networks. Taking the matching feature of power generation equipment and energy storage charging demand as an example, it is obtained by calculating the difference sequence of power generation and energy storage charging demand. The state of charge and discharge power constraint feature of the energy storage device is determined by the relationship between the current state of charge and the safe operating range (usually 20%-80%). The load-energy response feature is obtained by comparing the load change rate with the adjustable capacity of the energy storage.

[0062] In the output layer, the time difference algorithm is used for online learning. In implementation, the discount factor γ is set to 0.9 to balance the current decision and future impact. The learning formula based on the time difference algorithm is:

[0063] Q(s t ,h t ,a t )=r t +γQ(s t ,h t ,a t )

[0064] where s t+1 is the state vector, h t+1 is the dynamic coupling relationship feature, and at r is the power allocation scheme t γ is the discount factor that balances the current decision and future impact, Q(s t ,h t ,a t ) is the value assessment of the power allocation scheme in the current state, Q(s t+1 ,h t+1 ,a t+1 ) is the value assessment of the next state.

[0065] In implementing the ε-greedy strategy, the initial ε value is set to 0.9, indicating a 90% probability of choosing the highest value scheme and a 10% probability of random exploration. As the system runs and learns, the ε value gradually decreases and eventually stabilizes at 0.1 to maintain the system's adaptability to environmental changes.

[0066] The network finally outputs three types of power allocation values: the power generation allocation value of the power generation equipment, ranging from 0 to rated power; the charge and discharge power allocation value of the energy storage device, with charging being negative and discharging being positive; and the power distribution allocation value of the power load, which needs to be ensured not to exceed the rated power of the load.

[0067] In practical applications, the AI decision piece can perform decision calculation every 5 minutes, considering the physical constraints such as device start-stop delay and climbing rate during the calculation process to ensure that the output power allocation scheme is practically operable. For the calculated power allocation scheme, the system will convert it into specific device control instructions through the control execution piece.

[0068] The control execution piece includes an instruction conversion unit and a control execution unit, responsible for converting the power allocation scheme output by the AI decision piece into device control instructions and ensuring their execution. As shown in Figure 4 the control instruction execution process includes six main steps: first, receive the power allocation scheme output by the AI decision piece; then, according to the device type, perform instruction conversion; then, according to the communication protocol, perform message encapsulation; send control instructions through the gateway base module; collect device running status for feedback confirmation; finally, when execution is abnormal, resend the instruction until execution is successful or the upper limit of resending is reached.

[0069] In the implementation of the instruction conversion unit, the system converts the instruction according to the device type. For power generation devices, the system converts the set value according to the power value. For example, when the AI decision piece outputs a power value of 30 kW, the instruction conversion unit converts it into a standard power generation power instruction, which contains a power set value of 30 kW and a corresponding control word for controlling the output power of the power generation device. For energy storage devices, the system judges and converts the mode according to the positive and negative of the power value. For example, when the input power value is +20 kW, it is converted into a discharge power instruction, which contains a discharge mode flag and a power value of 20 kW; when the input power value is -20 kW, it is converted into a charging power instruction, which contains a charging mode flag and a power value of 20 kW. For power consumption loads, the power value is directly converted into a power consumption power limit instruction. For example, when the input power value is 15 kW, the converted instruction contains a power upper limit value of 15 kW and a corresponding limit flag bit.

[0070] The control execution unit encapsulates the converted control instruction according to the communication protocol requirements of the device after receiving it. Taking the Modbus protocol as an example, the encapsulated message contains: device address, function code, data area (containing the converted instruction content) and check value. For example, the message structure of the power setting instruction is:

[0071] [device address] [function code] [start register] [register number] [data length] [power value] [check code]

[0072] The control execution unit sends the encapsulated message to the corresponding device through the gateway basic module. At the same time, the system will collect the power, voltage and current running state of the device as feedback information, which is used to confirm the execution of the instruction. When the system detects that the instruction execution is abnormal, it will automatically resend the control instruction until it receives the correct execution state feedback or reaches the upper limit of the number of retransmissions.

[0073] In specific applications, the instruction conversion and execution process of each type of device strictly follows the above rules to ensure that the power distribution scheme of the AI decision piece can be accurately converted into executable control instructions for the device, and the execution of the instructions is confirmed through state feedback.

[0074] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions described in the foregoing embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An AI-based intelligent energy gateway intelligent scheduling and optimization control system, characterized in that, The system comprises an energy access piece, a data processing piece, an AI decision piece and a control execution piece; The energy access piece is configured to access and collect real-time operation data of power generation equipment, energy storage devices and power consumption loads supporting different communication protocols, and uniformly convert data of different protocols into a system standard data format comprising equipment identification, time stamp, operation data and state information; The data processing piece is configured to perform differential processing on the operation data in the standard data format by calculating the variation of adjacent sampling data, perform anomaly detection on data exceeding a preset threshold, and distribute valid data to the AI decision piece; The AI decision piece is configured to construct a load prediction model using a time sequence convolution network (TCN), and calculate an optimal power distribution scheme based on the prediction result using a deep reinforcement learning network (DQN); The control execution piece is configured to convert the power distribution scheme into equipment control instructions and execute the instructions while monitoring the instruction execution status; The deep reinforcement learning network (DQN) comprises an input layer, a hidden layer and an output layer: The input layer is configured to construct an input vector comprising voltage, current and power operation data of the power generation equipment, energy storage devices and power consumption loads transmitted by the data processing piece, and predicted load values, load variation trend features and load fluctuation ranges output by the time sequence convolution network (TCN); The hidden layer is configured to construct a state vector from the input vector, process the state vector through neural network weights and activation functions, automatically extract dynamic coupling relationship features between the equipment, wherein the dynamic coupling relationship features comprise matching features of power generation equipment and energy storage device charging demand, used to represent the surplus power generation available for energy storage; constraint features of energy storage device state of charge and discharge power, used to represent the available capacity and maximum charge and discharge power of the energy storage device; and response features of power consumption load variation and energy storage device charge and discharge adjustment, used to represent the demand amount of load fluctuation on energy storage adjustment; The output layer is configured to guide power distribution decisions using the dynamic coupling relationship features, calculate an optimal power distribution scheme based on the deep reinforcement learning network (DQN), perform nonlinear transformation calculation of the state vector and dynamic coupling relationship features using a neural network to obtain a value evaluation value, and perform online learning of the neural network using a time difference algorithm, wherein the time difference algorithm formula is: Q(s t ,h t ,a t ) = r t + γQ(s t+1 ,h t+1 ,a t+1 ) where s t is the state vector, h t is the dynamic coupling relationship feature, a t is the power allocation scheme, r t is the system reward value of the current power allocation scheme, γ is a discount factor that weighs the current decision and future impact, Q(s t ,h t ,a t ) is the value evaluation value of the power allocation scheme under the current state, and Q(s t+1 ,h t+1 ,a t+1 ) is the value evaluation value of the next time state. Based on the value evaluation result, an ε-greedy strategy is used to select a power distribution scheme with the highest value evaluation value, and output power generation power distribution values of the power generation equipment, charge and discharge power distribution values of the energy storage device, and power distribution values of the power consumption load.

2. The AI-based intelligent energy gateway scheduling and optimization control system of claim 1, wherein The energy access piece comprises a gateway basic module, a protocol analysis module, a protocol conversion module and a data cache module: The gateway basic module is configured to establish and maintain a bidirectional communication connection with the power generation equipment, energy storage devices and power consumption load equipment, support multiple device communication protocols, and realize data collection and control instruction transmission; The protocol analysis module is configured to analyze device communication frames of different protocols and extract voltage, current and power operation data; The protocol conversion module is configured to convert the parsed device data into system standard data format, including a device identification field, a timestamp field, a running data field, and a state field, wherein the device identification field includes a device type and a device communication address, the timestamp field is a UNIX timestamp of a data collection time, the running data field includes a device voltage value, a current value, and a power value, and the state field includes a device running state and a communication state. The data caching module is configured to cache the converted standard format data locally. 3.The AI-based intelligent energy gateway scheduling and optimization control system of claim 2, wherein, The data processing slice includes a data difference unit, a data verification unit, and a data distribution unit. The data difference unit is configured to perform difference processing on the running data in the standard data format cached by the data caching module, set difference threshold values of voltage, current, and power according to device types, calculate the change amount of adjacent sampling data and compare it with the corresponding preset threshold value, and when the change amount exceeds the preset threshold value, transmit the latest data to the data verification unit for processing. The data verification unit is configured to check the data transmitted by the data difference unit, determine whether the voltage, current, and power data exceed the rated value of the device, and determine the data as valid data when the voltage, current, and power data are within the rated value of the device and the data transmission is normal. The data distribution unit is configured to transmit the verified data to the AI decision slice in real time. 4.The AI-based intelligent energy gateway scheduling and optimization control system of claim 1, wherein, The time sequence convolution network TCN constructs a load prediction model, which specifically includes: establishing a historical operation database to store historical voltage, current, and power data of the power generation equipment, energy storage device, and power load; based on the historical operation data, constructing a time sequence convolution network TCN containing multiple layers of cavity convolution layers, and realizing long-term dependency analysis of historical data through an incremental cavity rate and a residual connection structure; outputting a predicted load value, a load change trend feature, and a load fluctuation range. 5.The AI-based intelligent energy gateway scheduling and optimization control system of claim 2, wherein, The control execution slice includes an instruction conversion unit and a control execution unit. The instruction conversion unit is configured to convert the power value output by the AI decision slice into communication control instructions for various devices: for power generation equipment, the power value is converted into a power generation power instruction conforming to its communication protocol; for energy storage devices, positive power is converted into a discharge power instruction, and negative power is converted into a charging power instruction; for power loads, the power value is converted into a power consumption power limit instruction; The control execution unit is configured to encapsulate the converted control instructions into messages according to the communication protocol requirements of each device and send them to the corresponding device through the gateway foundation module, confirm the execution status of the control instructions based on the collected device power, voltage, and current running state feedback information, and resend the instructions when an instruction execution anomaly is detected.

Images