An artificial intelligence-based new energy power station operation and maintenance method and system

Abstract

The application discloses a new energy power station operation and maintenance method and system based on artificial intelligence, and relates to the photovoltaic operation and maintenance management field. The method comprises the following steps: collecting the output power of each group of strings and the module temperature on the side of the current combiner box, calculating the power mean and variance, combining the temperature to generate node characteristics, and obtaining a full-station environment vector; constructing a graph structure by taking the group string as a node and the electrical connection as an edge, inputting the graph structure and the environment vector into a graph attention network, and obtaining a mismatch score of each group of strings; constructing a decision model by taking the mismatch score and the number of optimizers as inputs, solving an optimal compensation scheme, and issuing a compensation instruction to a corresponding DC / DC optimizer to realize dynamic mismatch compensation. Therefore, the adaptive ability of the photovoltaic power station under complex environment is improved, and the intelligent operation and maintenance level is improved.

Description

Technical Field

[0001] The present application relates to the field of photovoltaic operation and maintenance management technology, and in particular to an artificial intelligence-based new energy power station operation and maintenance method and system. Background Art

[0002] With the large-scale development of photovoltaic power plants, booster stations play a key role in aggregating the DC power from individual strings, boosting the voltage, and connecting them to the grid. However, complex factors such as varying component aging rates, partial shading, surface contamination, and uneven sunlight distribution can cause mismatches between strings on the DC side, leading to reduced power generation efficiency across the entire power plant. Currently, passive bypass diode technology or centralized DC / DC optimizers are the primary solutions to address string mismatch.

[0003] In passive bypass diode technology, a bypass diode is connected in parallel to each PV module. When the module's output power is significantly lower than that of other modules in the same string, the bypass diode conducts, cutting off the module's impact on the entire series current. However, bypass diodes are often accompanied by additional power losses and only isolate the faulty module without reconstructing the string's IV characteristic curve. Therefore, they cannot eliminate the impact of mismatches on system-level MPPT (Maximum Power Point Tracking).

[0004] In the booster station, each string is equipped with an independent DC / DC converter, independently adjusting the voltage-current characteristics of each string to achieve string-level MPPT optimization and achieve output close to the maximum power point. Although each string-level DC / DC optimizer can independently adjust the voltage-current curve based on electrical characteristics, its compensation decisions are still based solely on electrical measurement information, ignoring the on-site environment and operational conditions. For example, in situations such as sudden changes in light intensity, surface contamination, or shading of the component, relying solely on the current-voltage curve cannot accurately identify the root cause of the mismatch, resulting in compensation actions that are prone to target deviation or delayed response. Summary of the Invention

[0005] The present application provides an artificial intelligence-based new energy power station operation and maintenance method, system, storage medium, computer program product and electronic equipment to at least solve the problems of inaccurate detection of photovoltaic string mismatch, delayed response and unreasonable resource allocation in the current related technologies.

[0006] In the first aspect, the embodiment of the present application provides an artificial intelligence-based new energy power station operation and maintenance method, comprising: collecting the output power and module temperature values ​​of each photovoltaic string connected to the combiner box on the combiner box side, extracting the output power mean and output power variance of each photovoltaic string and generating a corresponding initial feature vector in combination with the module temperature value, and generating a global environment vector according to the environmental irradiance of the entire station and the environmental temperature of the entire station; constructing an undirected graph structure with each photovoltaic string as a node and the electrical connection relationship between each photovoltaic string as an edge; the node feature of each node in the undirected graph structure is defined according to the initial feature vector of the corresponding photovoltaic string; inputting the undirected graph structure and the global environment vector into a graph attention network, and L After the iterative fusion of the multi-head attention layers, L The layer output features are linearly mapped and ReLU activated to obtain the string mismatch score of each photovoltaic string; L represents the total number of graph attention layers in the graph attention network; constructs a compensation decision model with the string mismatch score of each photovoltaic string and the number of DC / DC optimizers as input; solves the compensation decision model to obtain the optimal compensation solution ;in, and Respectively represent The optimal target power and activation flag decision value corresponding to the PV string; according to the optimal compensation plan Generate compensation instructions and send them to the DC / DC optimizers corresponding to each photovoltaic string to implement dynamic compensation for string mismatch.

[0007] In the second aspect, the embodiment of the present application provides an artificial intelligence-based new energy power station operation and maintenance system, including: a multi-source feature acquisition unit, which is used to collect the output power and module temperature values ​​of each photovoltaic string connected to the combiner box on the combiner box side, extract the output power mean and output power variance of each photovoltaic string and generate a corresponding initial feature vector in combination with the module temperature value, and generate a global environment vector based on the environmental irradiance of the entire station and the environmental temperature of the entire station; a graph structure construction unit, which is used to construct an undirected graph structure with each photovoltaic string as a node and the electrical connection relationship between each photovoltaic string as an edge; the node feature of each node in the undirected graph structure is defined according to the initial feature vector of the corresponding photovoltaic string; a node feature update unit, which is used to input the undirected graph structure and the global environment vector into a graph attention network, and L After the iterative fusion of the multi-head attention layers, L The layer output features are linearly mapped and ReLU activated to obtain the string mismatch score of each photovoltaic string: LRepresents the total number of graph attention layers in the graph attention network; the compensation decision modeling unit is used to construct a compensation decision model based on the string mismatch score of each photovoltaic string and the number of DC / DC optimizers as input; the compensation decision solving unit is used to solve the compensation decision model to obtain the optimal compensation solution ;in, and Respectively represent The optimal target power and activation flag decision value corresponding to each photovoltaic string; the string mismatch dynamic compensation unit is used to calculate the optimal compensation solution. Generate compensation instructions and send them to the DC / DC optimizers corresponding to each photovoltaic string to implement dynamic compensation for string mismatch.

[0008] In a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the steps of the artificial intelligence-based new energy power station operation and maintenance method of any embodiment of the present application.

[0009] In a fourth aspect, an embodiment of the present application provides a storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, the steps of the artificial intelligence-based new energy power station operation and maintenance method of any embodiment of the present application are implemented.

[0010] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the artificial intelligence-based new energy power station operation and maintenance method of any embodiment of the present application.

[0011] The artificial intelligence-based new energy power station operation and maintenance method and system provided in this application can produce at least the following technical effects:

[0012] (1) Through multi-dimensional information fusion and intelligent optimization decision-making, a complete dynamic closed loop from mismatch signal capture to compensation execution is established. The power statistical characteristics and temperature data of each connected photovoltaic string are collected on the combiner box side. Combined with the whole station environmental monitoring, the graph structure constructed by electrical connections and the multi-head attention network are used to deeply mine the association between strings. Then, through optimized scheduling under resource constraints, the real-time accurate identification and rapid dynamic compensation of string mismatch are achieved, which improves the system-level maximum power tracking efficiency and the level of operation and maintenance automation.

[0013] (2) By collecting the output power mean, variance, and module temperature information of each photovoltaic string on the junction box side, and combining the global environmental irradiance and global environmental temperature to construct a global environmental vector, an undirected graph structure is used to characterize the connection relationship of each string, and the above features are input into the graph attention network for multi-level information fusion processing, thereby generating a robust mismatch degree score, achieving global perception and root cause location of the string mismatch degree. Compared with the current related technology that only detects based on voltage-current characteristics, this technical solution can effectively distinguish different types of mismatch problems such as component aging, shading, pollution, and uneven illumination, significantly improving the accuracy and pertinence of mismatch detection.

[0014] (3) A compensation decision model is constructed based on the string mismatch score and DC / DC optimizer resource constraints. By minimizing the system power loss and optimizing the resource activation cost, an optimal compensation strategy is constructed. The mismatch state is quickly identified and the optimal compensation strategy is generated. The string characteristics are dynamically adjusted by issuing compensation instructions, achieving a global optimal allocation of optimizer resources. This effectively optimizes the resource waste or unreasonable resource allocation problems of uncompensated mismatched strings in current related technologies, significantly improves the utilization efficiency of compensation resources, and ensures the maximization of overall power generation efficiency.

[0015] This technical solution accurately detects and dynamically compensates for string mismatches in PV strings on the combiner box side, effectively reducing power generation losses caused by string mismatches and improving the overall power generation efficiency of the PV power station. Furthermore, by rationally allocating and controlling the resources of each optimizer, it avoids energy waste caused by redundant compensation or uncompensated mismatched strings. It can quickly identify mismatches and generate optimal compensation strategies. By issuing compensation instructions, it dynamically adjusts string characteristics, significantly improving the PV power station's adaptability in complex environments, reducing the frequency of manual intervention, and enhancing intelligent operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0017] Figure 1 A flowchart illustrating an example of an artificial intelligence-based new energy power station operation and maintenance method according to an embodiment of the present application is shown;

[0018] Figure 2 A schematic diagram of a topological structure showing an example of a layout of power generation equipment of a new energy power station;

[0019] Figure 3 A schematic diagram showing an example of a current-voltage characteristic curve of a photovoltaic string;

[0020] Figure 4 An operational flow chart illustrating an example of constructing an undirected graph structure according to an embodiment of the present application is shown;

[0021] Figure 5 An operational flow chart illustrating an example of iteratively solving a compensation decision model using a three-stage alternating direction multiplier method according to an embodiment of the present application is shown;

[0022] Figure 6 A structural block diagram of an example of an artificial intelligence-based new energy power station operation and maintenance system according to an embodiment of the present application is shown;

[0023] Figure 7 This is a schematic diagram of the hardware structure of an electronic device for executing an artificial intelligence-based new energy power station operation and maintenance method provided in another embodiment of the present application. DETAILED DESCRIPTION

[0024] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0025] Figure 1 A flowchart of an example of an artificial intelligence-based new energy power station operation and maintenance method according to an embodiment of the present application is shown.

[0026] The execution entity of the method of the embodiment of the present application can be any controller or processor with computing or processing capabilities. Specifically, it can be implemented by a power station operation and maintenance platform or an edge ECU (Energy Control Unit) distributed across each combiner box. Through multi-dimensional intelligent feature extraction, graph network association analysis, and resource optimization and scheduling, it not only achieves high-precision identification and rapid compensation of PV string mismatches, but also significantly improves maximum power tracking efficiency, reduces manual operation and maintenance costs, and enhances the reliability and intelligence of booster station operations.

[0027] Figure 2 A topological structure diagram showing an example of the layout of power generation equipment in a new energy power station.

[0028] like Figure 2As shown, each PV string is first connected in parallel to its corresponding combiner box. The inverter then centrally inverts and boosts the voltage before connecting it to the utility grid. Specifically, PV strings 211, 213, and 215 are connected in parallel to combiner box 221 via their respective DC / DC optimizers 21a, 21c, and 21e, respectively. PV strings 212, 214, and 216 are connected in parallel to combiner box 222 via DC / DC optimizers 21b, 21d, and 21f, respectively. Combiner boxes 221 and 222 combine the DC currents from their respective parallel strings and feed them into inverter 240, where they are inverted and boosted before being connected to the grid for power generation.

[0029] To monitor power generation mismatches between strings in real time and perform dynamic compensation when necessary, edge ECUs (Edge ECU 231 and Edge ECU 232) are deployed on each combiner box. Specifically, each edge ECU continuously collects multi-source data from the multiple PV strings it connects to. Through edge computing, it identifies string mismatches between different PV strings within the combiner box and makes power compensation decisions. It then issues power compensation commands to the corresponding DC / DC optimizers, achieving dynamic compensation through precise adjustment of the power of each string.

[0030] In some examples, it can be integrated into an electronic device or terminal through software, hardware, or a combination of software and hardware, and the type of terminal or electronic device can be diverse, such as a mobile phone, tablet computer, or desktop computer, etc.

[0031] like Figure 1 As shown, in step S110, multi-source sensor data of each photovoltaic string connected to the combiner box is collected on the combiner box side, and the initial feature vector of each photovoltaic string is extracted, and a global environment vector is generated according to the global environmental irradiance and the global environmental temperature.

[0032] Specifically, the output power and module temperature values ​​of each photovoltaic string connected to the combiner box are collected on the combiner box side, the output power mean and output power variance of each photovoltaic string are extracted, and the corresponding initial feature vector is generated in combination with the module temperature value.

[0033] It should be noted that the fundamental cause of PV string mismatch is asynchronous fluctuations in output power between different PV strings under the same junction box due to factors such as shadowing, component aging, and thermal attenuation differences. Therefore, relying solely on a single electrical parameter source (such as voltage or current) cannot accurately reflect the true degree of mismatch, nor can it distinguish between environmental fluctuations and component anomalies. In the embodiments of the present application, by collecting multi-source sensor data and synchronously collecting and preprocessing power, module temperature, and ambient meteorological indicators, "external disturbances" and "internal states" are quantified separately, allowing the extracted structure to reflect the string's own performance while eliminating initial features of environmental interference.

[0034] In some embodiments, a power sensor (for measuring output power) and a module surface temperature sensor for each connected photovoltaic string are installed on each junction box side, and the corresponding power sampling frequency and temperature sampling frequency are configured respectively. For example, a set of shunt resistors are connected in series on the DC busbar connected between the junction box and each photovoltaic string to measure the real-time output power of each photovoltaic string. In addition, a thermocouple (such as PT100) is fixed on the back or welding strip area of ​​the representative module of each string to measure the surface temperature in real time. In addition, a two-axis photovoltaic dedicated irradiance meter is installed at the central location of the power station or deployed in zones to measure the irradiance of the entire station. The installed ambient temperature and humidity sensors measure the ambient temperature of the entire station. .

[0035] All channels are uniformly divided into time windows with a step size of 100ms. During the sampling process, the latest data in each time slot is timestamped and stored in a circular queue in the edge ECU to ensure that subsequent channel data can correspond one to one.

[0036] In terms of initial feature calculation, for the output power mean , can be calculated in each 30-second time window:

[0037] , formula (1)

[0038] In the time window length (like Calculate the arithmetic mean within 5 minutes):

[0039] , Formula (2)

[0040] Output power variance Calculated in the same time window:

[0041] , Formula (3)

[0042] The variance value directly reflects the power fluctuation amplitude and provides quantitative support for mismatch fluctuation.

[0043] The module temperature value can take the latest smoothed value of the temperature signal within the time window as the representative temperature input , no more statistics are needed to ensure sensitivity to hot spot mutations.

[0044] Then, the three are spliced ​​together to obtain the initial feature vector .

[0045] In addition, the time series exponential smoothing of the whole station irradiance and the whole station ambient temperature is performed, and the latest smoothed value can be used as the whole station ambient irradiance and the whole station ambient temperature, and the two can be fused as the global environment vector.

[0046] In this embodiment, a 30-second time window sampling and filtering strategy is adopted, and through smoothing and synchronization processing, power and temperature jumps caused by sudden weather changes (such as cloud shadows and wind and dust) are quickly filtered out, ensuring that subsequent fusion focuses on the real mismatch signal.

[0047] Figure 3 The figure shows an effect diagram of an example of a current-voltage characteristic curve of a photovoltaic string.

[0048] like Figure 3 As shown, the light yellow and orange lines represent the current-voltage characteristic curves (i.e., IV curves) of the same PV string at 25°C and 45°C, respectively. As the module temperature increases, the open-circuit voltage decreases significantly (the right end of the curve shifts downward), while the short-circuit current increases slightly (the left end of the curve shifts upward). However, the voltage-current product at the maximum power point (the dot) is generally lower than the performance at lower temperatures. This characteristic shift is a direct reflection of the increased carrier recombination and voltage decay in PV materials at high temperatures.

[0049] In the embodiment of the present application, considering the sensitive influence of temperature on the IV curve and the position of the maximum power point, the module temperature is used as an important dimension for mismatch assessment. Compared with only output power or electrical quantity as the basis, the accuracy of mismatch identification can be effectively improved by fusing temperature features. In addition, when a hot spot appears in a string due to obstruction, local damage or poor contact of the welding strip, the local module temperature will be significantly higher than that of the surrounding modules, resulting in the current output of the string being limited, while the adjacent strings may maintain normal output due to the lower temperature. Therefore, the temperature difference can not only quickly reflect the abnormality inside or on the surface of the component, but also cause a current "restraint" effect in the same junction box, thereby exacerbating the mismatch between strings. Incorporating module temperature into the features helps the model to give priority to those nodes that have mismatches due to temperature anomalies during fusion, and supplement the internal fault signals that are difficult to capture by electrical quantities alone from a physical level.

[0050] In step S120, an undirected graph structure is constructed with each photovoltaic string as a node and the electrical connection relationship between the photovoltaic strings as an edge. The node feature of each node in the undirected graph structure is defined according to the initial feature vector of the corresponding photovoltaic string.

[0051] It should be noted that the electrical connection relationship between strings determines the coupling mode of their output power, that is, the power changes of strings connected in parallel with the same combiner box are often correlated. Specifically, the connection relationship between each string and its combiner box is used as the physical connection basis.i Road group string and j If the road group strings are located in the same junction box, draw an undirected edge between them.

[0052] It should be understood that the initialization method of the edge weight of each edge in the undirected graph structure can be diverse, such as initialization according to a preset value. In addition, the edge weight can also be set according to the correlation of the data trend statistical results.

[0053] Figure 4 An operational flowchart of an example of constructing an undirected graph structure according to an embodiment of the present application is shown.

[0054] like Figure 4 As shown, in step S410, an undirected edge is constructed for any pair of nodes in the node set corresponding to each photovoltaic string, forming a complete graph between the entire set of nodes. In step S420, for each undirected edge in the complete graph, the Pearson correlation coefficient is calculated based on the output power time series data of the two photovoltaic strings corresponding to the undirected edge within a preset time window, and the absolute value of the Pearson correlation coefficient is used as the initial weight of the undirected edge.

[0055] In some embodiments, the output power of each string is collected synchronously at a fixed rate in the edge ECU or combiner box side controller. , set the time window length (For example, 5 minutes), before each compensation scheduling, collect the most recent Power timing data within.

[0056] For every edge in the complete graph , respectively take the power sequence of the two strings at the same time point 、 , according to the Pearson correlation coefficient calculation formula:

[0057] , Formula (4)

[0058] Calculate Pearson correlation coefficient , and take the absolute value:

[0059] , Formula (5)

[0060] The initial weight of each edge Store the corresponding adjacency matrix position.

[0061] It should be noted that the Pearson correlation can reflect the degree of synchronous fluctuation of the output power of the two strings under the same meteorological and load conditions. A high correlation indicates that they may be jointly affected by local shading or radiation gradients, and the absolute value can effectively reflect the corresponding correlation strength, thereby dynamically capturing short-term dynamic effects such as shadow movement and cloud changes.

[0062] In step S430, for each node in the complete graph, the initial weights of all connected edges of the node are normalized by row so that the sum of the normalized weights of all outgoing edges of each node is 1, thereby updating the edge weights of each connected edge.

[0063] In some embodiments, for each node , take the initial weight set of all connected edges , calculate their sum

[0064] , Formula (6)

[0065] Set the final weight of each edge to:

[0066] , formula (7)

[0067] Then, the normalized weighted adjacency matrix is ​​stored in the edge ECU. The normalization operation prevents the power amplitude from directly affecting the attention, ensuring that all nodes in their neighborhoods are fairly distributed with attention.

[0068] It should be noted that if fixed initial edge weights are used (for example, uniformly set to 0 or other constants), it is impossible to distinguish the contributions of different neighbors to the mismatch assessment, resulting in the neglect of neighborhood differences and high noise in the aggregated information. In contrast, in the embodiments of the present application, initial edge weights are assigned based on the absolute value of the Pearson correlation coefficient of the power time series. Through dynamic correlation weighting, multi-source information on the environment and component status is supplemented at the physical level. It also provides a prior distribution of mathematical trend statistics, which helps to automatically focus on those strings with highly synchronized or synchronously reversed operating fluctuations during the node aggregation stage, thereby truly reflecting the synergistic effects of disturbances such as dynamic electrical coupling and shadowing.

[0069] In step S130 , the undirected graph structure and the global environment vector are input into the graph attention network to perform node feature aggregation, thereby obtaining the string mismatch score of each photovoltaic string.

[0070] It should be noted that the environmental parameters of the entire site can have an overall trend impact on all photovoltaic strings. By using graph structure modeling and electrical topology to link the strings, and then aggregating the node features through graph neural networks, each string can not only "see" its own status with adjacent strings, but also synchronously "perceive" the global environment, thereby obtaining a mismatch score that comprehensively considers local and global factors.

[0071] More specifically, across the entire power plant, irradiance and ambient temperature are the most critical external factors determining the output power of PV modules, and their impact on the performance of each string is uniform and synchronized. Using the global ambient irradiance and temperature to construct a global environmental vector distinguishes overall power fluctuations caused by weather changes (such as rapid cloud movement or fluctuations in sunlight intensity) from the local mismatch behavior of a single string. Therefore, injecting this global vector as contextual information into the graph attention network allows the model to automatically "deduct" equilibrium changes caused by environmental fluctuations when calculating attention weights between nodes, thereby focusing on true component differences and electrical coupling anomalies.

[0072] Specifically, L After the iterative fusion of the multi-head attention layers, L The layer output features are linearly mapped and ReLU activated to obtain the string mismatch score of each photovoltaic string:

[0073] , formula (8)

[0074] Where, Indicates the String mismatch score of the PV string, represents the linear mapping trainable weight vector, represents the vector transpose operation, represents the linear mapping bias term, L represents the total number of graph attention layers in the graph attention network, Indicates the PV strings passing through L The feature vector output by the layer graph attention layer after fusion.

[0075] During the attention fusion process, each node inputs its initial feature vector and the global environment vector, which is not included in the node's features, into the attention scoring function in the same frame. Using a multi-head attention mechanism with branched parallel computation, the model learns multiple association patterns at different "points of attention." Through feature iterations across multiple attention layers, the weighted features of neighboring nodes are fused with the node's own features layer by layer, gradually expanding the receptive field. Through linear mapping and ReLU activation, the fused features are uniformly mapped to a non-negative scalar score, which measures the degree of mismatch between the current cluster and the overall cluster.

[0076] Through the learnable attention mechanism of the Graph Attention Network (GAT), the edge weight of each edge is dynamically adjusted, allowing the model to adaptively focus on the neighboring nodes that contribute most to the string mismatch score. At the same time, the attention intensity is adjusted in combination with the global environmental context, so that the differentiated associations of different strings in terms of mismatch degree can be effectively reflected in the node aggregation process.

[0077] Specifically, in the case of string mismatch, if the output fluctuations of a particular string differ significantly from those of some of its neighbors (i.e., high dissimilarity), GAT automatically amplifies the edge weights of these dissimilarities through an attention mechanism, while devaluing irrelevant or highly similar edge weights. This ensures that each node ultimately aggregates information that best represents its "local mismatch degree" and "neighborhood comparison relationship." After multiple layers of iterative fusion, the node forms a high-order representation that combines local features and neighborhood differences. Finally, through linear mapping and nonlinear activation, this representation is compressed into a non-negative scalar, the string mismatch score, which quantifies the difference between the output of this string and the expected / neighboring output in the current environment and neighborhood.

[0078] In step S140 , a compensation decision model is constructed with the string mismatch score of each photovoltaic string and the number of DC / DC optimizers as input.

[0079] It should be noted that the string mismatch score represents the degree to which each string deviates from its neighboring strings relative to their maximum output capacity. For strings with severe mismatch or high contribution, a power optimization reference is provided to the corresponding DC / DC optimizer to compensate for the mismatch. Furthermore, overcompensation should be avoided to prevent system oscillation and maintain low optimizer resource consumption.

[0080] It should be understood that the compensation decision model can be diverse, such as a priority greedy allocation model, a particle swarm metaheuristic model, etc.

[0081] For example, in the priority greedy allocation model, instead of solving the global optimal binary mixed integer program, we first use the mismatch score Compensation range Directly calculate the marginal benefit of unit compensation for each string.

[0082] , formula (9)

[0083] In other words, when the When one optimizer is enabled and fully compensated, the maximum benefit that can be obtained is Through marginal revenue The prior sorting of the scalability of the scalability function allows the DC / DC optimizer to be activated first, which in turn minimizes the overall mismatch. However, the greedy strategy is an approximation and cannot guarantee global optimality.

[0084] Preferably, in some examples of the embodiments of the present application, by constructing an optimization model based on mismatch score and compensation cost, a set of optimal power settings and compensator activation schemes are solved under the premise of satisfying power constraints and the available number of optimizers to achieve dynamic compensation decisions.

[0085] , formula (10)

[0086] in, , .

[0087] Where, N Indicates the total number of photovoltaic strings. Indicates the number of DC / DC optimizers, For the The rated power of the PV strings. For the The target power after PV string compensation, Enable penalty coefficient for optimizer; For the The activation flag of the DC / DC optimizer corresponding to the PV string. Indicates the The DC / DC optimizer is enabled for the photovoltaic strings. Indicates the The DC / DC optimizer is not enabled for the PV strings.

[0088] In the compensation decision model, the compensation power of each string is Mismatch score Forming product terms , which intuitively reflects the “mismatch loss that can be eliminated by compensating 1 watt”; the additional penalty term The resource cost of enabling each DC / DC optimizer is quantified as a unified “price” to constrain it. And the total number of enabled optimizers does not exceed the number of available optimizers Under the premise that the model minimizes the sum of total loss and total resource overhead, the optimal selection of "which groups should be compensated and how much watt should be compensated" is directly derived. Thus, from the physical layer, the power loss distribution of each group and the DC / DC resource consumption of the combiner box are considered, and the artificial rules or greedy algorithm are no longer relied on. Instead, the model automatically outputs the globally optimal "start-stop + power" decision, ensuring that each compensation can achieve the maximum mismatch reduction effect with the least hardware activation, significantly enhancing the reliability of the operation and maintenance strategy.

[0089] In step S150, the compensation decision model is solved to obtain an optimal compensation scheme.

[0090] Here, the optimal compensation scheme can be represented as , and respectively represent the optimal target power and activation flag decision value corresponding to the i-th road photovoltaic group string.

[0091] The solving method of the compensation decision model can be diversified, for example, a greedy heuristic algorithm, a genetic algorithm or an ADMM (Alternating Direction Method of Multipliers) method can be used, which is not limited here.

[0092] For example, in the greedy heuristic algorithm, based on the ratio of "unit compensation benefit" to the activation cost, all group strings are sorted from high to low, and the optimizer is assigned in turn, and then the projection or local optimization of the activated road is performed. Although this strategy cannot guarantee global optimality, it is simple to implement and has extremely low computational overhead, which is suitable for cases with extremely high real-time requirements or large group string scale.

[0093] In step S160, compensation instructions are generated according to the optimal compensation scheme , and the generated compensation instructions are sent to the DC / DC optimizer corresponding to each road photovoltaic group string to implement group string mismatch dynamic compensation.

[0094] Exemplarily, the edge ECU sends the compensation instructions to the DC / DC optimizer corresponding to each group string through the industrial bus, so that the DC / DC optimizer adjusts the output power according to , and adjusts through the internal boost / buck switch, thereby completing the compensation execution.

[0095] ​This embodiment introduces multi-source information fusion and a graph attention network. First, environmental interference is filtered out at the timing level, and power, module temperature, and meteorological parameters are organically integrated to accurately extract component mismatch signals physically. Then, on a fully connected graph constructed using the electrical topology, the contribution of each neighbor to the mismatch score is adaptively allocated, accurately quantifying the degree of mismatch for each string. Finally, combining the score with an optimization model based on optimizer resource constraints, a global optimal scheduling of power compensation requirements and optimizer activation costs is achieved, and this scheduling is distributed via a low-latency bus. This enables precise dynamic mismatch compensation and promotes intelligent and efficient operation and maintenance of photovoltaic power stations.

[0096] Regarding the description of the attention network in the figure in the embodiment of the present application, it can be fine-tuned on the framework of the general GAT to meet the actual needs of photovoltaic string mismatch assessment.

[0097] Specifically, during the offline training phase of the graph attention network, a large-scale dataset of time-series graph snapshots can be constructed. For each sampling moment, the current site-wide irradiance, ambient temperature, and the mean and variance of each string's output power, as well as module temperature, are used as node features. A normalized weighted adjacency matrix is ​​calculated based on the complete graph structure, forming a complete graph structure. Simultaneously, actual mismatch indicators at the corresponding moment are extracted from historical SCADA (Supervisory Control and Data Acquisition) records—for example, string mismatch magnitude calculated based on bypass diode triggering logs or inverter-side unbalanced current—to serve as labels for supervised learning. To ensure the model's robustness to varying weather conditions, seasons, and fault types, the training set must cover a variety of typical mismatch scenarios, such as sunny shadows, local hotspots, and module degradation. Samples are then balanced or weighted to ensure the model is not overly biased towards common scenarios while also identifying rare but critical extreme mismatches.

[0098] In constructing the graph attention model, a multi-layer GAT architecture is used, with 2-3 layers configured, each with 4-8 attention heads. Dropout (random dropout) of 0.2-0.5 is added after each layer to prevent overfitting, while LayerNorm (layer normalization) ensures training stability. The loss function is primarily based on mean squared error (MSE), with additional weighting applied to high mismatch errors to highlight key nodes—that is, a higher gradient weight is given to larger labels.

[0099] Finally, the verified model is exported into a lightweight inference format and deployed to the edge ECU for low-latency inference. Offline retraining is performed regularly (e.g. monthly) using the latest collected data to account for distribution drift caused by component aging or site topology changes. Combined with the electrical and environmental characteristics of the PV strings, this ensures that the final mismatch scoring model is superior to traditional static MPPT or bypass isolation solutions in terms of accuracy, real-time performance, and maintainability.

[0100] Regarding the details of calculating the string mismatch score of each photovoltaic string using the graph neural network, in some examples of the embodiments of the present application, the node initial feature matrix is ​​represented as:

[0101] , formula (11)

[0102] in, Indicates the The initial eigenvector corresponding to the photovoltaic string, and Respectively represent the first The arithmetic mean and standard deviation of the output power of the photovoltaic strings. Indicates the The surface module temperature of the PV panels used in the PV strings.

[0103] The global environment vector is expressed as:

[0104] , formula (12)

[0105] Where, represents the global environment vector, represents the ambient irradiance of the entire station, and Indicates the ambient temperature of the entire station.

[0106] Concatenate the node initial feature matrix and the global environment vector in the feature dimension into the initial fusion matrix :

[0107] , Formula (13)

[0108] It should be noted that string mismatch—the deviation between actual output and expected or adjacent string output—depends not only on the string's own electrical state (such as output power fluctuations and module temperature differences), but is also significantly affected by current / voltage coupling between adjacent strings in the same enclosure and overall meteorological conditions (irradiance and ambient temperature). Therefore, a preliminary fusion of node characteristics with the global environmental context is performed. Node characteristics reflect both local mismatch signals and the overall trends brought about by environmental changes, helping to distinguish environmental disturbances from component anomalies.

[0109] For the Layer attention heads, and the projection matrix is ​​expressed as , the attention parameter vector is expressed as , then calculate the original attention score of the neighbor nodes to the node.

[0110] The multi-head attention mechanism learns multiple neighbor information fusion modes in parallel for the same node in each layer. Each head calculates the original attention score in different projection subspaces based on the interaction between the node itself and the neighbor features and the prior edge weights.

[0111] , formula (14)

[0112] Where, Denotes the original attention score, which is expressed in Tier In the attention head, the neighbor node corresponding to the PV string pair The original influence of the photovoltaic string in the feature aggregation; and Respectively represent After layer iteration The node characteristic vector of the PV string and the The node feature vector of the photovoltaic string.

[0113] The normalized edge weight matrix of the re-fused node is used to update the original attention score:

[0114] , formula (15)

[0115] Where, Indicates the Tier In the attention head, for PV string node to the first The attention energy value with physical prior weighting of the photovoltaic string node; is the matrix element in the normalized edge weight matrix, indicating the PV string node to its neighbor The normalized edge weights of PV string nodes.

[0116] It should be noted that there may be differences in the current restraint effect of parallel groups in the same combiner box. By assigning initial edge weights based on the absolute value of the power time sequence Pearson correlation coefficient, mapping the synchronization / anti-synchronization information in real operation to the initial weights of the edges, distinguishing the contribution of different neighbors to the mismatch evaluation, and providing differentiated edge weight prior in the graph structure for different real-time correlation. By mapping the real-time correlation degree of different groups to the edge weight, the graph attention network can automatically identify and preferentially focus on those neighbors that are physically strongly coupled or have the most obvious synchronization mismatch when calculating the attention score of the target node.

[0117] For example, when the temperature of a certain string rises sharply due to local hot spots or poor solder strip contact, its power curve fluctuation characteristics will be highly correlated with those of other neighbors fluctuating due to environmental shading, light gradient, etc. At this time, the correlation coefficient is large, and the initial attention score also increases; on the contrary, when two strings differ greatly in environment and component state, the correlation coefficient is low, and the attention score is also suppressed, which usually means that the power fluctuations of the two are not caused by environmental factors (such as shading or temperature gradient) that can be compensated by adjusting the compensator power, but more likely reflect that one of the strings has a component defect or electrical fault, and relying solely on power compensation cannot effectively restore its state, and waste of compensation resources should be avoided. Therefore, after the normalized edge weight matrix of the node is fused to update the original attention score, more attention can be allocated to those nodes that really need priority compensation.

[0118] The attention weight is obtained by softmax normalization of all neighbor nodes of the node:

[0119] Equation (16)

[0120] In the formula, represents the normalized attention weight of the first layer of the first head, which is used to measure the relative importance of the first string to the first string in the channel; represents the set of all edges connected to the node of the first photovoltaic string, represents the placeholder index of any neighbor node directly connected to the node of the first photovoltaic string.

[0121] Here, the Softmax function is used to normalize all neighbors of the same node to obtain the final attention weight , which ensures that the sum of all is equal to 1 for a certain string , thereby forming a weight distribution with probability distribution properties.

[0122] Perform neighbor node feature aggregation on the node according to the attention weight:

[0123] , formula (17)

[0124] Where, Indicates the Tier Head to head The summary vector of a node in a group string represents the high-order features absorbed from all its neighbors and weighted transformed.

[0125] The first The features output by all attention heads of the layer are concatenated and activated to obtain the first Layer node characteristics:

[0126] , formula (18)

[0127] Where, Indicates the Tier The updated feature vector of the PV string, Indicates the The total number of attention heads in the layer.

[0128] Here, in the All in the layer After the outputs of the attention heads are spliced ​​together, the different neighborhood patterns captured by each attention sub-channel can be weighted and integrated through ReLU nonlinear activation to form the first High-order feature vectors of layer nodes. Multi-head concatenation not only greatly enriches the expressive power of node representations but also provides a wider feature subspace for the next layer, enabling the network to concurrently focus on multiple different mismatch triggering modes (such as hotspots, occlusions, and temperature fluctuations), thereby gradually refining more abstract and comprehensive mismatch representations through continuous iteration. ReLU activation further introduces nonlinear mapping capabilities, preserving positive responses while avoiding negative interference.

[0129] For the last layer L The node features output by the layer are linearly mapped and ReLU activated to obtain the string mismatch score of each photovoltaic string:

[0130] , formula (19)

[0131] Here, after multi-head and multi-layer attention weighted fusion, each node obtains a high-dimensional feature vector, which contains the multi-scale mismatch information of the node in the overall topology and environmental context. This high-dimensional vector is mapped to one dimension, and the non-negative part is retained by ReLU, thus completing the mapping from high-order representation to mismatch score.

[0132] Through the embodiments of the present application, a mismatch scoring method based on multi-source information fusion and weighted graph attention mechanism is proposed: by taking the power statistics, component temperature and environmental variables of each string as node features, and constructing a complete graph with the power time series correlation as the edge weight prior, the graph attention network can adaptively measure and aggregate multi-dimensional mismatch signals from different neighbors. The model is no longer limited to traditional data that relies only on voltage or current, but can simultaneously capture the dynamic coupling effects caused by temperature difference hotspots and local occlusion, thereby achieving high-precision quantification of the degree of string mismatch; at the same time, the multi-head parallel attention mechanism ensures that the network can pay attention to multiple mismatch trigger modes at the same time, avoiding misjudgment or information loss caused by the dominance of a single feature, and improving the recognition accuracy of string mismatch.

[0133] Regarding the solution details of the compensation decision model, in some examples of the embodiments of the present application, the compensation decision model can be iteratively solved by adopting a three-stage alternating direction multiplier method to obtain the optimal compensation solution. .

[0134] Figure 5 An operational flow chart illustrating an example of iteratively solving a compensation decision model using a three-stage alternating direction multiplier method according to an embodiment of the present application is shown.

[0135] like Figure 5 As shown, in step S510, in the first stage, the solution is obtained by independent optimization ( P , x ) sub-problem.

[0136] During the first phase of implementation, power compensation With binary enabled In global constraints The coupling problem under the , transforming the original coupled mixed integer programming into an ADMM alternating optimization framework. Specifically, the auxiliary variables Responsible for carrying the constraint of "the upper limit M of the total number of available optimizers", and the original binary variable In the first stage of independent optimization, the {0,1} value is completely retained. In this way, the complex coupling problem originally subject to the dual constraints of power demand and total resources is decomposed into “independent optimization of each string”. ” and “to all The two major sub-problems of "projection" are solved, thus achieving the decoupling and parallelization of the problem structure.

[0137] Specifically, in the In the round iteration, based on the auxiliary variables obtained in the previous round iteration and Lagrange multipliers , solve each photovoltaic string separately:

[0138] , formula (20)

[0139] Where, represents the augmented Lagrangian penalty coefficient, Indicates the During the round of iteration, The Lagrange multiplier corresponding to the photovoltaic string; Indicates the During the round of iteration, Auxiliary variables corresponding to the PV strings.

[0140] Therefore, the decomposition strategy ensures that: on the one hand, it can still be solved efficiently while retaining the binary decision, without On the other hand, the global resource constraints are only placed on the The global coupling is greatly reduced, allowing all sub-problems to run in parallel, greatly improving the edge device's ability to make real-time compensation decisions for more than 100 strings.

[0141] for If the update , the loss term follows increases and decreases linearly, the optimal ;like , then force the setting ; Indicates the At the beginning of the iteration, The enable flag variable corresponding to each PV string.

[0142] It should be noted that in the first stage Px sub-problem, for each string, if the flag is enabled , then the loss term Follow Linearly decreasing. Based on this property, the model directly converts Take the maximum value ;like , then the injunction To avoid invalid compensation, the power update can be divided into two extreme cases: full compensation when enabled and zero power output when disabled.

[0143] Thus, the mismatch score is fully utilized The significance of the power compensation indicator is that for enabled strings, increasing the power directly to the maximum available level can maximize the reduction of mismatch losses. For disabled strings, setting the power to 0 avoids compensating unnecessary nodes, saving optimizer resources. This ensures maximum compensation for each channel and makes overall compensation more targeted and efficient.

[0144] for The update of and The full cost in both cases, the cost function is given by And definition.

[0145] exist When the corresponding full cost is .

[0146] exist When the corresponding full cost is .

[0147] The one with the smaller complete cost value is selected as .

[0148] Here, for each string activation flag, The model enumerates the complete cost values ​​in the two cases of 0 and 1, and selects the one that minimizes the cost according to the size of the two values. , which not only takes into account the mismatch reduction benefits brought by enabling, but also is compatible with the auxiliary variables and multipliers There are two major factors in the consistency penalty.

[0149] By directly enumerating and comparing the costs, the system does not need to Relaxation and then backtracking to the binary solution can select the most appropriate start and stop decision in this round of iteration, avoiding the common search tree and backtracking calculation in mixed integer programming. This allows the algorithm to maintain the accuracy of each round when the group size is large. The updated linear time complexity significantly improves the edge real-time scheduling performance and ensures the binary executableness of the final compensation action.

[0150] In step S520, in the second phase, the auxiliary variables are solved by global resource constraint projection Sub-problem.

[0151] Fix the solution of the first stage , and the temporary auxiliary variable Projection to Satisfaction In the convex set:

[0152] Temporary auxiliary variables clipped to [0, 1]: ; represents the intermediate value after being forced to be within the interval [0, 1]. .

[0153] If , then ,

[0154] Otherwise, use the bisection method to obtain the offset , and let .

[0155] Here, the global constraint "enable the optimizer total number does not exceed " and the interval constraint of each node auxiliary variable are organically divided into two steps: first, a "clipping" operation is performed on each temporary calculation value, and then only when the sum of the clipped exceeds , a bisection search is triggered to accurately calculate an offset , so that all simultaneously satisfy the single-node interval constraint and the global sum is strictly equal to . Thus, when the sum after clipping itself does not exceed the limit, the is directly taken as the final output, saving computing overhead; when the sum after clipping exceeds the limit and needs to be adjusted, the offset obtained based on the bisection method not only ensures that all linear or nonlinear constraint conditions are not violated, but also quickly completes the projection, greatly improving the solving efficiency of the edge ECU and ensuring convergence stability.

[0156] In step S530, in the third phase, the Lagrange multiplier is updated.

[0157] , equation (21)

[0158] In the equation, represents the Lagrange multiplier after the first round of update, so that in the next round of iteration of the first round, it continues to promote the reduction of the gap between and .

[0159] By cumulatively updating the gap between the binary variable and the auxiliary variable , the multiplier term in the next round of iteration is equivalent to a "pulling force", allowing the subsequent and More inclined to take the same value, round by round to push them to be consistent. Thus, by multiplier update, ADMM iteration can guarantee that the global resource constraints and local decision after several iterations perfect match, no oscillation or no solution convergence problem.

[0160] It should be noted that, compared to the ordinary Lagrange relaxation method in the absence of augmented terms may appear "dual saddle point" dilemma, augmented multiplier ensures the convergence and stability of the algorithm, while in the actual deployment to achieve the With The gap is less than the set threshold to terminate the iteration, significantly shortening the overall solution time.

[0161] In the embodiments of the present application, when applying the three-stage alternating direction multiplier method (ADMM) to solve the photovoltaic string mismatch compensation problem, the main difference from the traditional ADMM algorithm is: first, the binary constraint of the enable flag is always retained , and the two values are compared by direct enumeration in the subproblem, avoiding continuous relaxation and complex integerization steps of ; second, in the update of the power variable , the end point extreme value can be directly taken due to its linear structure, so that the originally coupled two-dimensional optimization is decomposed into two-step parallel update of each string, greatly reducing the calculation amount of the subproblem. Third, in handling the global compensator number constraint, through the truncation and ordering projection algorithm of the auxiliary variable , the global constraint projection of is quickly completed, saving the complex quadratic programming solution that may occur in the general ADMM. Thus, through these improvements, the three-stage alternating direction multiplier method can achieve higher decision execution efficiency.

[0162] Further, by repeatedly performing the above three-stage operation iteration until The gap between is less than the preset convergence threshold , it is determined that the iteration converges, a set of converged variables and are obtained, and the optimal compensation scheme is obtained through the corresponding string parameter setting:

[0163] , equation (22)

[0164] In the formula, represents the target power value of the th photovoltaic string after iteration convergence, represents the enable flag value of the th photovoltaic string after iteration convergence, represents the auxiliary variable value of the th photovoltaic string after iteration convergence.

[0165] Here, no additional integerization is required after convergence, which not only ensures the binary executable of the solution but also avoids the time overhead and error caused by backtracking after relaxation. It can be efficiently distributed to each road optimizer through the industrial bus, accurately reducing string mismatch and significantly improving the real-time, reliability and economy of system compensation decisions.

[0166] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of combined actions, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application. In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0167] Figure 6 A structural block diagram of an example of an artificial intelligence-based new energy power station operation and maintenance system according to an embodiment of the present application is shown.

[0168] like Figure 6 As shown, the artificial intelligence-based new energy power station operation and maintenance system 600 includes a multi-source feature acquisition unit 610, a graph structure construction unit 620, a node feature update unit 630, a compensation decision modeling unit 640, a compensation decision solving unit 650 and a string mismatch dynamic compensation unit 660.

[0169] The multi-source feature acquisition unit 610 is used to collect the output power and module temperature values ​​of each photovoltaic string connected to the combiner box on the combiner box side, extract the output power mean and output power variance of each photovoltaic string and generate the corresponding initial feature vector in combination with the module temperature value, and generate a global environment vector based on the global environmental irradiance and global environmental temperature.

[0170] The graph structure construction unit 620 is used to construct an undirected graph structure with each photovoltaic string as a node and the electrical connection relationship between the photovoltaic strings as an edge; the node feature of each node in the undirected graph structure is defined according to the initial feature vector of the corresponding photovoltaic string.

[0171] The node feature updating unit 630 is used to input the undirected graph structure and the global environment vector into the graph attention network. L After the iterative fusion of the multi-head attention layers, LThe layer output features are linearly mapped and ReLU activated to obtain the string mismatch scores of the PV strings. L denotes the total number of graph attention layers in the graph attention network.

[0172] The compensation decision modeling unit 640 is configured to construct a compensation decision model with the string mismatch scores of the PV strings and the number of DC / DC optimizers as inputs.

[0173] The compensation decision solving unit 650 is configured to solve the compensation decision model to obtain an optimal compensation scheme ; wherein, and denote the optimal target power and the enabled flag decision value of the i-th PV string, respectively.

[0174] The string mismatch dynamic compensation unit 660 is configured to generate compensation instructions according to the optimal compensation scheme and deliver the generated compensation instructions to the DC / DC optimizers corresponding to the PV strings to implement the string mismatch dynamic compensation.

[0175] In some embodiments, the present application provides a non-volatile computer readable storage medium, wherein the storage medium stores one or more programs including execution instructions, which can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to perform the steps of any of the above-mentioned artificial intelligence-based new energy power station operation and maintenance methods.

[0176] In some embodiments, the present application also provides a computer program product, which includes a computer program stored on a non-volatile computer readable storage medium, and the computer program includes program instructions, which, when executed by a computer, cause the computer to perform the steps of any of the above-mentioned artificial intelligence-based new energy power station operation and maintenance methods.

[0177] In some embodiments, the present application also provides an electronic device, which includes at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the steps of the artificial intelligence-based new energy power station operation and maintenance method.

[0178] Figure 7 is a hardware structure schematic diagram of an electronic device for executing the artificial intelligence-based new energy power station operation and maintenance method provided by another embodiment of the present application, as shown in Figure 7 The device includes: ​

[0179] One or more processors 710 and memory 720, Figure 7 A processor 710 is taken as an example.

[0180] The device for executing the artificial intelligence-based new energy power station operation and maintenance method may further include: an input device 730 and an output device 740 .

[0181] The processor 710, the memory 720, the input device 730 and the output device 740 may be connected via a bus or other means. Figure 7 The bus connection is taken as an example.

[0182] Memory 720, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the artificial intelligence-based new energy power station operation and maintenance method in the embodiments of this application. Processor 710 executes the non-volatile software programs, instructions, and modules stored in memory 720 to execute various server functional applications and data processing, thereby implementing the artificial intelligence-based new energy power station operation and maintenance method in the aforementioned method embodiment.

[0183] The memory 720 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device, etc. In addition, the memory 720 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 720 may optionally include a memory remotely located relative to the processor 710, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0184] The input device 730 may receive input digital or character information and generate signals related to user settings and function control of the electronic device. The output device 740 may include a display device such as a display screen.

[0185] The one or more modules are stored in the memory 720, and when executed by the one or more processors 710, the artificial intelligence-based new energy power station operation and maintenance method in any of the above method embodiments is executed.

[0186] The above-mentioned product can execute the method provided in the embodiment of this application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of this application.

[0187] The electronic devices of the embodiments of the present application exist in various forms, including but not limited to:

[0188] (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and their primary purpose is to provide voice and data communications. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0189] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers and have computing and processing capabilities, and generally also have mobile Internet access. These terminals include PDAs, MIDs, and UMPCs.

[0190] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players, handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0191] (4) Other onboard electronic devices with data interaction functions, such as onboard computer devices installed in vehicles.

[0192] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0193] Through the description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a general hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the relevant technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0194] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A new energy power station operation and maintenance method based on artificial intelligence, characterized in that: The method comprises: At the combiner box side, the output power and module temperature values ​​of each PV string connected to the combiner box are collected. The output power mean and output power variance of each PV string are extracted and combined with the module temperature value to generate the corresponding initial feature vector. The global environment vector is also generated based on the overall station ambient irradiance and overall station ambient temperature. An undirected graph structure is constructed using each photovoltaic string as a node and the electrical connection relationship between the photovoltaic strings as an edge; the node feature of each node in the undirected graph structure is defined according to the initial feature vector of the corresponding photovoltaic string; The undirected graph structure and the global environment vector are input into the graph attention network. L After the iterative fusion of the multi-head attention layers, L The layer output features are linearly mapped and ReLU activated to obtain the string mismatch score of each photovoltaic string; L Represents the total number of graph attention layers in the graph attention network; A compensation decision model is constructed using the string mismatch score of each PV string and the number of DC / DC optimizers as input; Solve the compensation decision model to obtain the optimal compensation solution ;in, and Respectively represent The optimal target power and activation flag decision value corresponding to each PV string; According to the optimal compensation plan Generate compensation instructions and send them to the DC / DC optimizer corresponding to each photovoltaic string to implement dynamic string mismatch compensation; The string mismatch score is expressed as follows: ， Where, Indicates the String mismatch score of the PV string, represents the linear mapping trainable weight vector, represents the vector transpose operation, represents the linear mapping bias term, Indicates the PV strings passing through L The feature vector output after the layer graph attention layer fusion; The compensation decision model is: ， in, , ; Where, N Indicates the total number of photovoltaic strings. Indicates the number of DC / DC optimizers, For the The rated power of the PV strings. For the The target power after PV string compensation, Enable penalty coefficient for optimizer; For the The activation flag of the DC / DC optimizer corresponding to the PV string. Indicates the The DC / DC optimizer is enabled for the photovoltaic strings. Indicates the The DC / DC optimizer is not enabled for the PV strings.

2. The method according to claim 1, characterized in that The undirected graph structure is constructed by taking each photovoltaic string as a node and the electrical connection relationship between the photovoltaic strings as an edge, including: In the node set corresponding to each photovoltaic string, an undirected edge is constructed for any pair of nodes to form a complete graph between the entire set of nodes; For each undirected edge in the complete graph, a Pearson correlation coefficient is calculated based on the output power time series data of the two photovoltaic strings corresponding to the undirected edge within a preset time window, and the absolute value of the Pearson correlation coefficient is used as the initial weight of the undirected edge; For each node in the complete graph, the initial weights of all connected undirected edges of the node are normalized by row so that the sum of the normalized weights of all connected undirected edges of each node is 1, so as to update the edge weights of each connected undirected edge.

3. The method according to claim 2, characterized in that The undirected graph structure and the global environment vector are input into the graph attention network, L After the iterative fusion of the multi-head attention layers, L The layer output features are linearly mapped and ReLU activated to obtain the string mismatch score of each photovoltaic string, including: The node initial feature matrix is ​​expressed as: ， in, Indicates the The initial eigenvector corresponding to the photovoltaic string, and Respectively represent the first The arithmetic mean and standard deviation of the output power of the photovoltaic strings. Indicates the The surface module temperature of the photovoltaic modules used in the photovoltaic strings; The global environment vector is expressed as: ， Where, represents the global environment vector, represents the ambient irradiance of the entire station, and Indicates the ambient temperature of the entire station; Concatenate the node initial feature matrix and the global environment vector in the feature dimension into the initial fusion matrix : ， For the Layer attention heads, and the projection matrix is ​​expressed as , the attention parameter vector is expressed as , then calculate the original attention score of the neighbor node to the node: ， Where, Denotes the original attention score, which is expressed in Tier In the attention head, the neighbor node corresponding to the PV string pair The original influence of the photovoltaic string in the feature aggregation; and Respectively represent After layer iteration The node characteristic vector of the PV string and the The node feature vector of the photovoltaic string; The normalized edge weight matrix of the re-fused node is used to update the original attention score: ， Where, Indicates the Tier In the attention head, for PV string node to the first The attention energy value with physical prior weighting of the photovoltaic string node; is the matrix element in the normalized edge weight matrix, indicating the PV string node to its neighbor Normalized edge weights of PV string nodes; The attention weight is obtained by performing softmax normalization on all neighbor nodes of the node: ， Where, Indicates the Tier The normalized attention weight of the head is used to measure the Road group string pair The relative importance of the road group string in the channel; Indicates the The set of all edges connected to the photovoltaic string nodes. Indicates the The placeholder index of any neighbor node directly connected to the PV string node; Perform neighbor node feature aggregation on the node according to the attention weight: ， Where, Indicates the Tier Head to head The summary vector of the path group string node represents the high-order features absorbed from all neighbors and weighted transformed; The first The features output by all attention heads of the layer are concatenated and activated to obtain the first Layer node characteristics: ， Where, Indicates the Tier The updated feature vector of the PV string, Indicates the The total number of attention heads in the layer; For the last layer L The node features output by the layer are linearly mapped and ReLU activated to obtain the string mismatch score of each photovoltaic string: 。 4. The method according to claim 3, characterized in that Solving the compensation decision model to obtain the optimal compensation solution ,include: The three-stage alternating direction multiplier method is used to iteratively solve the compensation decision model to obtain the optimal compensation solution. ; In the first stage, the solution is solved by independent optimization ( P , x ) Sub-question: In the In the round iteration, based on the auxiliary variables obtained in the previous round iteration and Lagrange multipliers , solve each photovoltaic string separately: ， Where, represents the augmented Lagrangian penalty coefficient, Indicates the During the round of iteration, The Lagrange multiplier corresponding to the photovoltaic string; Indicates the During the round of iteration, Auxiliary variables corresponding to the PV strings; for If the update , the loss term follows Increases and decreases linearly, the optimal ;like , then force the setting ; Indicates the At the beginning of the iteration, The enable flag variable corresponding to each PV string; for The update of and The full cost in both cases, the cost function is given by And definition; exist When the corresponding full cost is ; exist When the corresponding full cost is ; The one with the smaller complete cost value is selected as ; In the second stage, auxiliary variables are solved by global resource constraint projection Sub-questions: Fix the solution of the first stage , and the temporary auxiliary variable Projection to Satisfaction In the convex set: Temporary auxiliary variables Clip to [0,1]: ; Indicates that The intermediate value after being constrained to the interval [0,1]; like ,but , Otherwise, use the binary method to find the offset , and then ; In the third phase, the Lagrange multipliers are updated: ， Where, Indicates in The Lagrange multiplier after the round of update makes the Continue to promote the reduction of and the gap; Repeat the above three-stage operation iteration until and The gap is less than the preset convergence threshold , then the iteration is judged to converge and a set of convergent variables is obtained and And get the optimal compensation solution by setting the corresponding string parameters: ， Where, After the iteration converges, The target power value corresponding to each photovoltaic string is After the iteration converges, The enable flag value corresponding to the PV string. After the iteration converges, The auxiliary variable value corresponding to each PV string.

5. An artificial intelligence-based new energy power station operation and maintenance system, characterized in that: The system comprises: The multi-source feature acquisition unit is used to collect the output power and module temperature values ​​of each photovoltaic string connected to the combiner box on the combiner box side, extract the output power mean and output power variance of each photovoltaic string, and generate the corresponding initial feature vector based on the module temperature value. In addition, the global environment vector is generated based on the global environmental irradiance and global environmental temperature. A graph structure construction unit is configured to construct an undirected graph structure using each photovoltaic string as a node and the electrical connection relationship between the photovoltaic strings as an edge; the node feature of each node in the undirected graph structure is defined based on the initial feature vector of the corresponding photovoltaic string; A node feature updating unit is used to input the undirected graph structure and the global environment vector into the graph attention network, L After the iterative fusion of the multi-head attention layers, L The layer output features are linearly mapped and ReLU activated to obtain the string mismatch score of each photovoltaic string: L Represents the total number of graph attention layers in the graph attention network; A compensation decision modeling unit is used to construct a compensation decision model using the string mismatch score of each photovoltaic string and the number of DC / DC optimizers as input; Compensation decision solving unit, used to solve the compensation decision model to obtain the optimal compensation solution ;in, and Respectively represent The optimal target power and activation flag decision value corresponding to each PV string; String mismatch dynamic compensation unit is used to calculate the optimal compensation solution Generate compensation instructions and send them to the DC / DC optimizer corresponding to each photovoltaic string to implement dynamic string mismatch compensation; The string mismatch score is expressed as follows: ， Where, Indicates the String mismatch score of the PV string, represents the linear mapping trainable weight vector, represents the vector transpose operation, represents the linear mapping bias term, Indicates the PV strings passing through L The feature vector output after the layer graph attention layer fusion; The compensation decision model is: ， in, , ; Where, N Indicates the total number of photovoltaic strings. Indicates the number of DC / DC optimizers, For the The rated power of the PV strings. For the The target power after PV string compensation, Enable penalty coefficient for optimizer; For the The activation flag of the DC / DC optimizer corresponding to the PV string. Indicates the The DC / DC optimizer is enabled for the photovoltaic strings. Indicates the The DC / DC optimizer is not enabled for the PV strings.

Images