Converter harmonic suppression method and system fusing multi-modal data

Abstract

The embodiment of the application relates to the technical field of harmonic suppression of a converter, and provides a harmonic suppression method and system of a converter fusing multi-modal data, which comprehensively depict the dynamic operation state of the converter by fusing current, voltage, temperature and other multi-modal data streams in real time, and performs timing risk prediction based on the same, so as to identify the potential harmonic over-standard risk in advance, accurately locate the problem source through a causal attribution mechanism when the harmonic over-standard risk is predicted, and generate optimal avoidance actions by using a dynamic model to deduce, finally analyze new control parameters, and use the new control parameters to perform smooth scheduling on the converter. In this way, the robustness and self-adaptive ability of the converter system to unknown dynamic working conditions such as grid voltage drop and load mutation are significantly enhanced, so as to comprehensively guarantee the power quality and operation reliability of the converter in the full working condition range.

Description

Technical Field

[0001] This invention relates to the field of converter harmonic suppression technology, and in particular to a converter harmonic suppression method and system that integrates multi-mode data. Background Technology

[0002] In modern power systems and industrial production, converters, as core equipment for power conversion, are widely used in many fields such as new energy power generation, motor drives, and rail transit traction. However, the switching operation of power devices inside the converter inevitably injects harmonic currents and voltages into the power grid, causing power quality problems. These harmonics not only cause additional losses in power equipment such as transformers and motors, leading to overheating and reduced efficiency, but also disrupt communication systems, even triggering local resonance, threatening the safe and stable operation of the power grid, and under certain conditions, causing malfunctions or damage to sensitive electronic equipment. Therefore, developing efficient and reliable methods to suppress converter harmonics is of paramount importance for ensuring the stable operation of the power system, improving power quality, and extending equipment lifespan.

[0003] Currently, most traditional harmonic suppression methods rely on the design and optimization of models and parameters established under finite, steady-state or quasi-steady-state operating conditions. These models typically focus on capturing the correlations between variables and can achieve good suppression effects under ideal or pre-set operating environments. However, real-world converter systems are complex dynamic physical systems with rapidly changing operating conditions. When the system encounters sudden and drastic load changes, such as motor stall or a deep voltage drop in the grid, or parameter drift caused by component aging, the generalization ability of existing models to unknown dynamic conditions becomes significantly insufficient, and the prediction accuracy drops sharply. In the process of complex dynamic changes in operating conditions, these static, correlation-based models often misjudge, resulting in poor harmonic compensation effects. They not only fail to effectively suppress harmonics but may even cause more serious power quality problems or equipment risks due to incorrect response strategies. This lack of robustness to unknown dynamic conditions has become a key obstacle to current converter harmonic suppression technology. Summary of the Invention

[0004] The present invention aims to solve at least one of the problems existing in the prior art, and to provide a converter harmonic suppression method and system that integrates multimodal data.

[0005] One aspect of the present invention provides a converter harmonic suppression method that integrates multimodal data, comprising: Acquire raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream; Feature engineering is performed on the original three-phase current data stream, original three-phase voltage data stream, original DC bus voltage data stream, original temperature data stream, and original load command asynchronous data stream to obtain the time series of state feature vectors; Time series risk prediction based on state feature vectors to obtain trigger signals; In response to the trigger signal being true, causal attribution and candidate action generation are performed on the current state feature vector to obtain a candidate action set; The candidate action set and the current state feature vector are input into the trained dynamic model to obtain the optimal avoidance action; The optimal avoidance action is analyzed to obtain new control parameters, and the converter is smoothly scheduled using the new control parameters.

[0006] Another aspect of the present invention provides a converter harmonic suppression system that integrates multimodal data, comprising: The data acquisition module is used to acquire raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream; The feature engineering module is used to perform feature engineering on the original three-phase current data stream, the original three-phase voltage data stream, the original DC bus voltage data stream, the original temperature data stream, and the original load command asynchronous data stream to obtain the time series of state feature vectors; The temporal risk prediction module is used to predict temporal risks based on the time series of state feature vectors in order to obtain trigger signals. The causal attribution and candidate action generation module is used to perform causal attribution and candidate action generation on the current state feature vector in response to the trigger signal being true, so as to obtain a set of candidate actions. The optimal avoidance action dynamic generation module is used to input the candidate action set and the current state feature vector into the trained dynamic model to obtain the optimal avoidance action. The action analysis and smooth scheduling module is used to analyze the optimal avoidance action to obtain new control parameters, and then use the new control parameters to perform smooth scheduling of the converter.

[0007] Compared with existing technologies, this invention provides a converter harmonic suppression method and system that integrates multimodal data. By fusing multimodal data streams such as current, voltage, and temperature in real time, it comprehensively depicts the dynamic operating state of the converter and performs time-series risk prediction based on this data. This allows for early identification of potential harmonic exceedance risks. When a harmonic exceedance risk is predicted, a causal attribution mechanism is used to accurately locate the root cause of the problem, and a dynamic model is used to generate the optimal avoidance action. This is ultimately resolved into new control parameters, which are then used to smoothly schedule the converter. This approach significantly enhances the robustness and adaptability of the converter system to unknown dynamic conditions such as grid voltage dips and load surges, thereby comprehensively ensuring the power quality and operational reliability of the converter across the entire operating range. Attached Figure Description

[0008] One or more embodiments are illustrated by way of example with the corresponding pictures in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0009] Figure 1 A flowchart of a converter harmonic suppression method that integrates multimodal data according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the data flow in the converter harmonic suppression method that integrates multimodal data according to an embodiment of the present invention; Figure 3 This is a block diagram of a converter harmonic suppression system that integrates multimodal data according to an embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the various embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0011] As indicated in the specification and claims of this invention, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0012] While this invention makes various references to certain modules in systems according to embodiments of the invention, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0013] This invention uses flowcharts to illustrate the operations performed by the system according to embodiments of the invention. It should be understood that the preceding or following operations are not necessarily performed in precise order. Instead, various steps can be processed in reverse order or simultaneously, as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0014] In the technical solution of this invention, a converter harmonic suppression method that integrates multimodal data is proposed. Figure 1 This is a flowchart of a converter harmonic suppression method that integrates multimodal data according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the data flow in a converter harmonic suppression method that integrates multimodal data according to an embodiment of the present invention. (In conjunction with...) Figure 1 and Figure 2 According to an embodiment of the present invention, a converter harmonic suppression method integrating multimodal data includes the following steps: S1, acquiring raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream; S2, performing feature engineering on the raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream to obtain a time series of state feature vectors; S3, performing time-series risk prediction based on the time series of state feature vectors to obtain a trigger signal; S4, responding to the trigger signal being true, performing causal attribution and candidate action generation on the current state feature vector to obtain a candidate action set; S5, inputting the candidate action set and the current state feature vector into a trained dynamic model to obtain the optimal avoidance action; S6, performing action analysis on the optimal avoidance action to obtain new control parameters, and using the new control parameters to smoothly schedule the converter.

[0015] Specifically, S1 involves acquiring the raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream. Traditional harmonic suppression methods typically rely on static models trained under limited, steady-state or quasi-steady-state operating conditions. This leads to a significant decrease in prediction accuracy when the converter system encounters sudden and drastic load changes or key parameter drifts, resulting in poor harmonic compensation or even erroneous judgments. Therefore, in the technical solution of this invention, by acquiring the raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream, the dynamic operating status information of the converter and its grid and load sides is comprehensively and in real-time captured. This provides a solid data foundation for subsequent feature engineering, timing risk prediction, causal attribution, and intelligent decision-making, ensuring an accurate understanding of the root causes of harmonic generation under complex and dynamic operating conditions and enabling timely and effective suppression decisions.

[0016] The raw three-phase current data stream refers to the real-time measurement sequence of three-phase currents collected from the AC output side of the converter or the grid side. This data is typically acquired using high-precision current sensors such as Hall effect sensors and current transformers. After analog-to-digital conversion (ADC), it forms a continuous digital data stream at a predetermined sampling frequency. It directly reflects the current injected by the converter into the grid or load and is the basis for evaluating harmonic distortion, load status, and system dynamic response. The raw three-phase voltage data stream refers to the real-time measurement sequence of three-phase power line voltages or three-phase voltages collected from the converter's connection point to the grid. This data is typically acquired using voltage sensors, voltage transformers, and other devices and digitized by an ADC. It provides instantaneous waveform information of the grid voltage and is a key basis for analyzing power quality events such as grid harmonic background, voltage dips, or transient fluctuations. The raw DC bus voltage data stream refers to the real-time measurement sequence of DC voltage acquired from the DC side bus of the converter. It is typically obtained through voltage divider resistors, voltage sensors, etc., and digitized by an ADC. Its fluctuations, especially ripple, directly reflect the operating status of the DC-side energy storage components and the stability of the converter's switching action, and are closely related to the generation and propagation mechanisms of harmonics. The raw temperature data stream refers to the real-time temperature sequence acquired from key components inside the converter, especially power device modules such as the heat sinks or module surfaces of insulated-gate bipolar transistors (IGBTs). This data is usually acquired through thermistors, thermocouples, or other temperature sensors. Given that the junction temperature of power devices has a significant impact on their reliability, lifespan, and electrical parameters such as switching losses, and that temperature changes indirectly affect the nonlinear characteristics and switching behavior of the devices, thus influencing harmonic generation, the raw temperature data stream is of great significance for assessing the health status of the converter and for causal attribution. The raw load command asynchronous data stream refers to the real-time control command sequence received by the converter control system, indicating information such as its output power or torque. It is an important reference for judging the current operating condition of the converter system, the driving cause, and predicting potential risks. All these raw data streams—raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream—are continuously and frequently acquired through appropriate sensors and data acquisition units, and transmitted to the processing unit, laying a solid foundation for subsequent feature extraction and analysis.

[0017] Specifically, in step S2, feature engineering is performed on the raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream to obtain a time series of state feature vectors. It should be understood that when facing the complex and ever-changing operating environment of a converter system, the acquired raw data streams often have high dimensionality, high frequency, heterogeneity, and may contain noise. Directly using this raw data for subsequent risk prediction and decision-making not only incurs a huge computational burden but may also reduce the accuracy and generalization ability of the model due to redundant information and noise interference. Therefore, in the technical solution of this invention, feature engineering is performed on the raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream to refine and abstract the multimodal raw data, transforming it into a series of physically meaningful key features that can effectively characterize the real-time state of the converter system and have predictive value. This process can unify raw, discrete, and potentially asynchronous data into synchronous, structured state feature vectors, thereby providing a standardized, efficient, and information-rich input for subsequent time-series risk prediction, causal attribution, and the generation of optimal avoidance actions.

[0018] The state feature vector refers to a high-dimensional data point at a specific moment, composed of multiple calculated and filtered key indicators. It comprehensively reflects various key parameters of the converter system's current operation. Specifically, the state feature vector includes the RMS current value, total harmonic distortion rate of the grid voltage, DC bus voltage ripple, load change rate, and estimated IGBT junction temperature. It is extracted and integrated from different raw data streams, together forming a comprehensive view of the system state.

[0019] In specific implementation, S2 firstly extracts and calculates the effective value of the current from the original three-phase current data stream. This effective value represents the equivalent value of the thermal effect of the AC current. Specifically, within a preset time window, such as one power frequency cycle or a sliding window, the instantaneous values ​​of the three-phase current are squared, then the average value of the squared values ​​is calculated, and finally the square root of the average value is taken to obtain the effective value of the current. Secondly, the total harmonic distortion rate of the grid voltage is calculated based on the original three-phase voltage data stream. The calculation of the total harmonic distortion rate of the grid voltage usually involves performing Fourier transform or similar spectral analysis on the instantaneous voltage waveform to separate the fundamental component and each harmonic component. Then, according to the standard definition, the square root of the sum of the squares of the effective values ​​of all non-fundamental harmonic components is calculated, and its ratio to the effective value of the fundamental voltage is calculated. This ratio quantifies the degree of distortion of the harmonic components in the grid voltage waveform relative to the fundamental wave, which is used as the total harmonic distortion rate of the grid voltage. Next, the DC bus voltage ripple is calculated based on the original DC bus voltage data stream. This step can be achieved by removing the DC component through high-pass filtering and then measuring the peak or RMS value of the AC component. For example, by measuring the maximum and minimum values ​​of the DC bus voltage within a specific time period and calculating the difference between them, the peak value of the ripple voltage can be approximated. This peak value reflects the smoothness of the DC power supply. Then, the load change rate is calculated based on the original load command asynchronous data stream. This can be achieved by differential processing of continuous load command values; that is, subtracting the load command from the previous time from the current load command, and then dividing the result by the time interval between the two time points. The final result is the load change rate, reflecting the rate of change of load demand per unit time. Furthermore, by combining the original temperature data stream with other electrical parameters, the junction temperature of the IGBT is estimated using a thermal model. IGBT junction temperature is difficult to measure directly. It is typically estimated in real-time using sensor-measured casing temperature, current flowing through the IGBT, voltage across the IGBT, and thermal resistance models of the power device such as Foster or Cauer networks. This estimation process may involve complex mathematical models and iterative calculations. The basic idea is that power loss causes junction temperature rise, and the temperature difference between the junction and casing temperatures is related to power loss and thermal resistance, thus providing temperature information for key nodes within the device. Synchronously integrating all the characteristic values ​​calculated at the same time point or time window—namely, the RMS current, total harmonic distortion of the grid voltage, DC bus voltage ripple, load change rate, and the estimated IGBT junction temperature—forms a complete state feature vector. The state feature vectors for all time points or time windows are continuously generated according to the sampling frequency, thus constructing a time series of state feature vectors reflecting the dynamic changes of the converter system. The state feature variables in the state feature vector are the RMS current, total harmonic distortion of the grid voltage, DC bus voltage ripple, load change rate, and estimated IGBT junction temperature.

[0020] Specifically, S3 involves performing time-series risk prediction based on the time series of state feature vectors to obtain trigger signals. During the operation of the converter system, the raw multimodal data collected in real time undergoes feature engineering processing to form a time series of state feature vectors that characterize the system state. However, simply possessing current or historical state information is insufficient to proactively address complex and ever-changing operating conditions. Therefore, in the technical solution of this invention, time-series risk prediction is performed based on the time series of state feature vectors to proactively identify potential harmonic exceedance risks by utilizing the dynamic patterns contained in the time series data, and to issue timely warning signals. This enables the compensation control to be predictive, allowing for advance adjustments to the compensation amount and ultimately improving dynamic response performance. This prediction mechanism avoids the lag of ex-post compensation in traditional methods, thereby initiating subsequent causal attribution and decision-making processes before the risk fully manifests.

[0021] The trigger signal is a control signal. When the predicted risk of harmonic exceedance reaches or exceeds the preset safety threshold, the trigger signal is set to true to initiate subsequent intervention measures.

[0022] In its specific implementation, S3 first involves performing state sequence temporal feature encoding on the time series of the state feature vector to obtain the final hidden state vector. That is, by performing state sequence temporal feature encoding on the time series of the state feature vector, advanced intelligent algorithms are used to extract deep-seated time dependencies and intermodal interaction information from historical and current state feature vector sequences. This refines the complex dynamic evolution process into a concise and predictive representation—the final hidden state vector—making subsequent risk assessments and decisions more forward-looking and accurate.

[0023] In this process, recurrent neural networks such as gated recurrent units (GRUs) are used as the basic encoder, combined with temporal attention, enabling the model to not only see the current state but also understand the trend of the state, which is crucial for dealing with dynamic operating conditions. Specifically, recurrent neural networks such as GRUs effectively control the flow of information at time steps through their internal gating mechanisms, including update gates and reset gates, thereby learning and memorizing the temporal dependencies in long sequences and alleviating the gradient vanishing problem of traditional recurrent neural networks. Through GRUs, the model can effectively learn the patterns and trends of various features such as the RMS value of current, total harmonic distortion of voltage, DC bus voltage ripple, load change rate, and estimated junction temperature of IGBTs over time from the time series of state feature vectors. Building upon this foundation, a hierarchical attention structure is further introduced: In the first layer, the model operates on the independent time series of each mode, namely the RMS current, total harmonic distortion of voltage, DC bus voltage ripple, load change rate, and estimated IGBT junction temperature, and applies an intra-mode attention mechanism to capture the key points of the dynamic evolution of that mode. This means that for each type of data stream, such as three-phase current data stream, the model evaluates the importance of data at different points in time within that mode to the overall trend and state changes of that mode, thereby finding the most representative time segments. For example, a brief but dramatic numerical fluctuation within a mode may be given a higher attention weight. Next, the second layer integrates and interacts with the core information extracted from each mode in the first layer across modes. Through this mechanism, the model can dynamically weigh and allocate the relative importance of data from different modes. For example, under specific operating conditions, the temperature mode, i.e., the change in the estimated IGBT junction temperature, may be more indicative of risk than the current mode, i.e., the change in the RMS current, and the model will accordingly give the temperature mode a higher degree of confidence. This hierarchical attention mechanism enables the model to understand the complex nonlinear coupling relationships between various physical quantities in the converter system more deeply and intuitively, enhancing the model's interpretability and ensuring that the final hidden state vector comprehensively and accurately reflects the dynamic characteristics of the converter system. After the fusion processing of the recurrent neural network encoding and the hierarchical attention mechanism, the time series of the state feature vector is transformed into a final hidden state vector with a fixed dimension. This final hidden state vector is a highly condensed abstract concept that contains key dynamic information, interrelationships, and implicit indications of future risks for all modes over a period of time. It serves as a direct and efficient input for subsequent risk probability regression and trigger signal generation.

[0024] Furthermore, risk probability regression and trigger signal generation are performed on the final hidden state vector to obtain the trigger signal. It should be understood that simply possessing a summary hidden state vector is insufficient to directly guide the behavior of the converter system, especially when proactive harmonic risk avoidance is required. Therefore, risk probability regression and trigger signal generation are performed on the final hidden state vector to transform the complex, high-dimensional hidden state information into an intuitive and understandable risk probability. Based on a preset risk tolerance, a clear trigger signal is then generated. In a specific example of this invention, if the predicted harmonic exceedance risk reaches a critical level, the subsequent causal attribution and action generation process is triggered, ensuring that the converter system can respond to potential risks in a timely and targeted manner, thereby greatly improving dynamic response performance.

[0025] In this process, firstly, the final hidden state vector undergoes nonlinear mapping and probability transformation to obtain the risk probability. Specifically, the final hidden state vector is input into a regression model specifically designed for risk assessment. This regression model incorporates a series of complex nonlinear mappings, such as using a Multilayer Perceptron (MLP) or other deep neural network structures, to learn the complex nonlinear relationship between the final hidden state vector and the harmonic exceedance risk. Through these nonlinear mappings, the regression model can extract patterns closely related to harmonic risk from the highly abstract hidden state and integrate them into a preliminary risk metric. This risk metric is further processed by a probability transformation function, such as the Sigmoid activation function. The Sigmoid activation function can compress and map any real value to the interval between 0 and 1, thereby standardizing the abstract risk metric into an intuitive and interpretable risk probability that explicitly represents the likelihood of a harmonic exceedance event occurring in the converter system within the current or future prediction window. For example, a risk probability close to 1 implies an extremely high harmonic exceedance risk, while a risk probability close to 0 indicates an extremely low harmonic exceedance risk.

[0026] Furthermore, a threshold decision is made based on a comparison between the risk probability and a preset risk threshold to obtain a trigger signal. That is, the risk probability is directly compared with a pre-set risk threshold. This preset risk threshold is a key parameter determined based on the safety requirements, cost-effectiveness, and tolerance for power quality in the actual application scenario, representing the lowest risk level that the converter system deems necessary for proactive intervention. If the calculated risk probability is greater than or equal to the preset risk threshold, it indicates that the currently predicted harmonic risk has reached or exceeded an acceptable range, and a trigger signal with a "true" status is generated. Conversely, if the risk probability is lower than the preset risk threshold, a trigger signal with a "false" status is generated, indicating that the converter system is currently operating relatively stably and there is no need to immediately initiate subsequent complex decision-making processes. The generation of trigger signals effectively filters out non-emergency situations, allowing computational resources to be focused on truly high-risk events that require attention.

[0027] Specifically, in step S4, in response to the trigger signal being true, causal attribution and candidate action generation are performed on the current state feature vector to obtain a candidate action set. It should be understood that in the practice of converter harmonic suppression, traditional causal attribution mechanisms have a core technical flaw: they ignore the universally existing synergistic causal effects, i.e., the special relationships of causal interaction, between variables in their analytical framework. They can only assess the independent main effect of a single controllable variable on harmonic exceedance, failing to effectively quantify the non-additive effects produced when multiple state variables or control variables are combined. This effect often far exceeds the simple superposition of the independent effects of each factor. In complex nonlinear systems like converters, harmonic problems are often not triggered by a single factor, but rather by multiple factors coupled and jointly excited under specific operating conditions. For example, in a specific application scenario, when the grid impedance is low, even if the load current is high, the harmonics may still be within the normal range; conversely, when the load current is small, even if the grid impedance is high, the harmonics may be small. However, when both high grid impedance and high load current occur simultaneously, the converter system is highly susceptible to resonance at a certain frequency, leading to a sharp amplification of that harmonic and severe exceedances. In this situation, the root cause of the harmonic exceedance is not a single factor—either grid impedance or load current—but rather the strong synergistic effect between the two. Traditional causal attribution mechanisms, by calculating the independent contribution of each variable separately, may find that neither is particularly prominent, thus incorrectly attributing the harmonic to other minor variables. Ultimately, this results in irrelevant candidate avoidance actions that fail to address the underlying harmonic problem, rendering the avoidance decision ineffective.

[0028] To address the aforementioned technical deficiencies, this invention provides a preferred solution that can simultaneously quantify the attribution framework of primary causal effects and co-causal effects, and identify the truly dominant causal pattern from it, thereby ensuring that the generated candidate actions can fundamentally solve the problem.

[0029] Specifically, causal attribution and candidate action generation are performed on the current state feature vector to obtain a candidate action set. This includes: First, based on the causal graph, main causal impact quantification is performed on the current state feature vector to obtain the main causal impact vector. That is, by establishing a reliable benchmark to measure the independent contribution of each factor, the independent influence of each state variable on the harmonic exceedance event is accurately quantified. Specifically, for the current state feature vector... By utilizing causal graphs learned offline and applying the do-operator for causal intervention, the effect of each state feature variable on harmonic exceedance events is calculated. The main causal impact In other words, based on the causal graph, the main causal impact vector is obtained by performing main causal impact quantization on the current state feature vector, including determining the main causal impact vector using the following formula: ; in, Represents the state feature vector at the current moment. The Middle State characteristic variables The current observation value, This indicates an event where harmonics exceed the standard. This indicates that a harmonic exceedance event has occurred. This indicates that a harmonic exceedance event has occurred. The baseline probability or prior probability, Indicates will The value is fixed as Time harmonic exceedance event The probability of occurrence is... Applying intervention Post-harmonic exceedance event The probability of occurrence; Indicates that in a given hour Harmonic exceedance events The primary causal impact, i.e., state characteristic variables Harmonic exceedance events The primary causal impact, i.e., the state characteristic variables Corresponding main causal impact This allows us to isolate the risk contribution driven by a single factor from a complex system state, providing a clear and quantitative benchmark for identifying more complex synergistic effects. It should be noted that... Indicates the number of the state characteristic variable, and , This represents the dimension of the state feature vector at the current moment. When the state feature vector includes the RMS current value, the total harmonic distortion rate of the grid voltage, the DC bus voltage ripple, the load change rate, and the estimated junction temperature of the IGBT, , The values ​​of 1, 2, 3, 4, and 5 correspond to five state characteristic variables: effective current value, total harmonic distortion rate of grid voltage, DC bus voltage ripple, load change rate, and estimated IGBT junction temperature, respectively. The principal causal impacts corresponding to all state characteristic variables constitute the principal causal impact vector.

[0030] Next, based on the causal graph, the cocausal impact quantization is performed on the current state feature vector to obtain the cocausal impact matrix. This step aims to quantify the additional causal influence beyond the sum of their independent effects when any two state variables act together, in order to capture and amplify the nonlinear interaction effect where 1+1>2. In this process, the original cocausal effect is first calculated. It intervenes in state characteristic variables through joint intervention. and state characteristic variables The total effect, minus the main causal shocks of each of the two state characteristic variables. The sum is obtained, and then a key, creative state coupling weight is introduced. This state coupling weight can dynamically amplify the synergistic effect when the two state feature variables are simultaneously in their extreme distribution regions. Finally, it can amplify the original synergistic effect. State-coupled weights Multiplying them together yields the final co-causal impact. In other words, based on the causal graph, the cocausal impact quantization is performed on the state feature vector at the current moment to obtain the cocausal impact matrix, including: determining the cocausal impact matrix using the following formula: ; ; ; in, This represents the state feature vector at the current moment. State characteristic variables The current observation value, This represents the state feature vector at the current moment. State characteristic variables The current observation value, This indicates an event where harmonics exceed the standard. This represents the baseline or prior probability of a harmonic exceedance event. Indicates will The value is fixed as At the same time The value is fixed as The probability of time harmonic exceedance events. express The corresponding primary causal impact, express The corresponding primary causal impact, express and Harmonic exceedance events The original synergistic effect, They represent , The value after Z-score standardization; This represents the scaling factor, a positive number used to control the sensitivity of the weighting function; Indicates the intervention condition is and State coupling weights during time; express and The co-causal impact, and also the first in the co-causal impact matrix Line 1 The elements of the column. This means that when or When at least one of them is close to its historical mean, its standardized value or Approaching zero, making It is also close to zero, indicating that the synergistic effect is not significant at this point. However, when and At the same time, when they deviate significantly from their mean, for example, when they are both high values ​​or one is high and the other is low. It will rapidly approach 1, thus greatly amplifying the original synergistic effect. In this way, the entire attribution model can, like an experienced engineer, keenly focus on dangerous scenarios where multiple factors simultaneously reach a critical state, leading to a problem's outbreak. This involves traversing... All possible values, i.e., 1 to D and All possible values, from 1 to D, can be determined according to... By obtaining all elements in the cocausal impact matrix, the cocausal impact matrix is ​​obtained.

[0031] Furthermore, based on the main causal impact vector, the co-causal impact matrix, and the action mapping table, dominant causal pattern identification and candidate action generation are performed to obtain a candidate action set. It should be understood that an effective decision must be based on a comprehensive comparison of all possible causes to identify the most critical driving factors. Therefore, the quantitative results of main effects and co-effects are further integrated to identify the true dominant causal pattern in the current state.

[0032] In this process, firstly, based on the main causal impact vector and the cocausal impact matrix, the current dominant factor is determined. That is, the main causal impact of all state characteristic variables is determined. Cocausal shocks corresponding to all state characteristic variables Compare them together, find the item with the largest absolute value, and identify it as the current dominant factor. In other words, based on the main causal impact vector and the cocausal impact matrix, the current dominant factor is determined, including by using the following formula: ; in, This indicates the current dominant factor. The th element in the current state feature vector State characteristic variables The corresponding primary causal impact, The first in the co-causal impact matrix Line 1 The elements of the column are also and Co-causal impact, This represents the state feature vector at the current moment. There are 1 state feature variables, where D is the dimension of the state feature vector at the current moment.

[0033] Next, based on the current dominant factor and action mapping table, the candidate action set is generated. That is, depending on the type of the current dominant factor, such as whether it is a single variable or a pair of variables, the corresponding candidate avoidance action set is retrieved from a predefined action mapping table as the candidate action set.

[0034] In this way, complex numerical calculation results can be transformed into a clear and interpretable diagnostic conclusion. In a specific example of this invention, if the current dominant factor is identified as a single variable, it indicates a relatively direct root cause of the problem; if the current dominant factor is identified as a pair of variables, it clearly points to the synergistic effect of the two variables as the root cause of the problem. Ultimately, a candidate action set consisting of one or more highly targeted candidate avoidance actions is generated. For example, if the current dominant factor is identified as... This combination, in which, Indicates the power grid impedance. The effective value of the current indicates that the generated action will be a systematic adjustment measure such as activating active damping control or fine-tuning the switching frequency to avoid the resonant point, rather than simply adjusting a single parameter. This ensures that the avoidance action is targeted and fundamentally improves the effectiveness of the decision.

[0035] In summary, the above-mentioned optimized scheme achieves a leap from simple single-factor causal attribution to the identification of complex main-effect-synergistic-effect composite causal patterns. Specifically, the optimized scheme constructs a harmonic risk attribution engine capable of deeply understanding the nonlinear coupling relationships within the converter system. This enables precise and targeted solutions, proposing truly effective avoidance strategies. It allows for the proactive identification of latent harmonic risks arising from the combined effects of multiple factors, which are easily overlooked by traditional methods. Furthermore, because this mechanism can more accurately pinpoint the root cause of harmonic problems, whether single or synergistic factors, the generated candidate avoidance actions are more targeted and effective, significantly improving the success rate of proactive harmonic avoidance and avoiding the waste of resources and potential risks caused by ineffective or erroneous avoidance operations. This allows the entire harmonic suppression process to evolve from a passive compensation mode to a more intelligent, reliable, and efficient proactive avoidance mode, ultimately comprehensively improving the power quality and operational reliability of the converter under complex dynamic conditions.

[0036] Specifically, in step S5, the candidate action set and the current state feature vector are input into the trained dynamic model to obtain the optimal avoidance action. In the preceding steps, a sophisticated causal attribution mechanism is used to identify the root cause of harmonic risk and generate a candidate action set containing multiple targeted solutions. However, simply generating a candidate action set is insufficient, because in complex converter systems, the actual effects of different actions under different operating conditions may vary significantly, and may even introduce new problems. To ensure the effectiveness, safety, and minimization of negative impacts of the suppression strategy, a mechanism is needed to evaluate these candidate actions and select the optimal solution. In the technical solution of this invention, inputting the candidate action set and the current state feature vector into the trained dynamic model enables accurate predictive evaluation of each potential avoidance action, i.e., simulating the possible consequences of the action under the current specific system state. This evaluation helps to identify those actions that can most effectively reduce harmonics while maintaining stable system operation and avoiding unnecessary costs.

[0037] The trained dynamic model is pre-built and trained using a large amount of historical or simulation data. Its training objective is to accurately simulate the behavior and response of the converter system under different states and after applying different control actions, especially its impact on key indicators such as harmonic levels, system stability, and operating efficiency. This dynamic model can be implemented using various technologies, such as physical models, data-driven models like deep learning, regression models, or reinforcement learning models, as long as its training objective is achieved.

[0038] In its implementation, S5 firstly, for each candidate action in the candidate action set, the trained dynamic model uses it and the current state feature vector as joint input. The core task of the dynamic model is to predict the behavior trajectory and final state of the converter system over a future period after executing the corresponding candidate action under the current state feature vector. During the prediction process, the dynamic model evaluates factors such as the expected harmonic suppression effect, the impact on other system parameters such as voltage stability and power factor, and potential transient responses or side effects. For each prediction result, its merits are quantified using a predefined utility function or reward function. This function comprehensively considers multiple performance indicators such as the degree of harmonic suppression, energy loss, equipment loss, impact on load, and system stability, calculating a quantified score for each candidate action. For example, a high-efficiency utility function will give a higher score to actions that significantly reduce harmonics and have the least impact on other performance aspects. Furthermore, after iteratively predicting and scoring all candidate actions in the candidate action set, the candidate action with the highest or best score is selected as the optimal avoidance action by comparing the utility scores of all candidate actions. This optimal avoidance action is the decision predicted by the dynamic model to achieve the best harmonic suppression effect and the best overall performance under the current system state.

[0039] Specifically, in step S6, the optimal avoidance action is analyzed to obtain new control parameters, and the converter is smoothly scheduled using these new control parameters. It should be understood that the actual operation and control of the converter relies on precise, real-time quantified control parameters, such as specific voltage and current reference values, switching frequency, and control gain. Directly applying an abstract action to a complex power electronic system is not only difficult to implement, but may also cause system oscillations, overshoot, or even equipment damage due to improper operation. Furthermore, sudden and significant changes in control parameters often lead to transient deterioration of the converter system, generating new harmonic interference or disrupting system stability. Therefore, to ensure that the generated candidate avoidance actions are more targeted and effective, the technical solution of this invention further analyzes the selected optimal avoidance action in detail, converting it into specific and executable new control parameters, and gradually and stably implementing parameter adjustments through a smooth scheduling mechanism to avoid unnecessary system shocks and ensure the smooth and safe operation of the converter.

[0040] In specific implementation, for parameter adjustment actions such as "fine-tuning the switching frequency," the process of determining the specific target frequency value in step S6 includes: Target value determination: Based on the current operating status, harmonic frequency analysis results, and system model, a specific switching frequency value that is far from the resonance point and meets the system performance requirements is calculated. For example, if the current harmonic is near a certain resonance frequency and the current switching frequency is an old value, a new switching frequency can be calculated according to a preset strategy. The principle is to add or subtract a frequency interval from the old value. The size of this frequency interval depends on the bandwidth of the resonance and the system response speed to ensure that the adjusted frequency can effectively avoid resonance. Parameter calibration: Ensure that the target value is within the physical limits of the converter and the numerical processing range of the controller.

[0041] For mode switching or module activation actions, such as "activating active damping control", implement the following steps: Instruction generation: Generates a binary signal or status variable that enables a specific functional module; Initial parameter setting: Set a set of initial operating parameters for the activated module, i.e., the specific functional module, such as the control gain of the active damper and the filtering time constant. These parameters can be obtained from the pre-configured configuration library or dynamically calculated according to the current converter system status to ensure smooth intervention.

[0042] Action analysis concretizes the optimal avoidance action into a set or series of new numerical or discrete control parameters, which the converter controller can directly understand and execute.

[0043] After obtaining the specific target control parameters, i.e., the new control parameters, these new control parameters are applied to the converter control system gradually rather than abruptly through smooth scheduling. This gradual adjustment is crucial for maintaining system stability. Specifically, the smooth scheduling process includes the following steps: First, if there is a significant difference between the new control parameters and the current actual parameters, a smooth transition trajectory from the current actual parameter value to the target parameter value corresponding to the new control parameters is generated. This can typically be achieved using methods such as ramp functions, S-curves, or other low-pass filters. For example, when adjusting the switching frequency from the current actual parameter value to the target parameter value, the scheduler may linearly increase or decrease the frequency at a fixed rate of change within a preset transition time. This rate of change can be obtained by calculating the ratio of the difference between the target frequency and the current frequency to the set transition time. Within each control cycle, the calculation principle of the new frequency command is to add or subtract the product of the rate of change and the time of a single control cycle based on the frequency command of the previous control cycle, until the current frequency reaches the target frequency value. It is worth noting that choosing an appropriate conversion time is crucial. It must ensure the smoothness of parameter changes to avoid transient shocks, while also being fast enough to handle real-time harmonic risks. Finally, the generated parameter commands for each moment are sent to the converter's underlying controller. During this process, the converter's response can be continuously monitored to ensure the smooth scheduling process proceeds as expected, and adjustments or rollbacks can be made promptly when anomalies are detected. Through smooth scheduling, new control parameters are gradually and safely integrated into the converter's operation, thereby effectively suppressing harmonics while maintaining system stability and reliability.

[0044] In summary, the converter harmonic suppression method integrating multimodal data according to embodiments of the present invention is explained. It comprehensively characterizes the dynamic operating state of the converter by real-time fusion of multimodal data streams such as current, voltage, and temperature, and performs time-series risk prediction based on this data. This allows for early identification of potential harmonic exceedance risks. When a harmonic exceedance risk is predicted, the root cause is accurately located through a causal attribution mechanism, and the optimal avoidance action is generated using a dynamic model. This is ultimately resolved into new control parameters, which are then used for smooth scheduling of the converter. This approach significantly enhances the robustness and adaptability of the converter system to unknown dynamic conditions such as grid voltage dips and load surges, thereby comprehensively ensuring the power quality and operational reliability of the converter across the entire operating range.

[0045] The present invention also provides a converter harmonic suppression system that integrates multimodal data.

[0046] Figure 3 This is a block diagram of a converter harmonic suppression system that integrates multimodal data according to an embodiment of the present invention. Figure 3As shown, the converter harmonic suppression system 300 integrating multi-modal data according to an embodiment of the present invention includes: a data acquisition module 310, used to acquire raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream; a feature engineering module 320, used to perform feature engineering on the raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream to obtain a time series of state feature vectors; and a time series risk prediction module 330, used to predict the state feature vectors based on the time series data stream. The time series of state feature vectors is used to predict the timing risk to obtain the trigger signal; the causal attribution and candidate action generation module 340 is used to generate a candidate action set by performing causal attribution and candidate action generation on the current state feature vector in response to the trigger signal being true; the optimal avoidance action dynamic generation module 350 is used to input the candidate action set and the current state feature vector into the trained dynamic model to obtain the optimal avoidance action; the action parsing and smooth scheduling module 360 ​​is used to perform action parsing on the optimal avoidance action to obtain new control parameters, and use the new control parameters to perform smooth scheduling of the converter.

[0047] The specific implementation method of the converter harmonic suppression system 300 that integrates multimodal data provided in this embodiment of the invention can be found in the converter harmonic suppression method that integrates multimodal data in this embodiment of the invention, and will not be repeated here.

[0048] The converter harmonic suppression system 300 integrating multimodal data according to embodiments of the present invention can be implemented in various wireless terminals, such as servers with converter harmonic suppression algorithms integrating multimodal data. In one possible implementation, the converter harmonic suppression system 300 integrating multimodal data according to embodiments of the present invention can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the converter harmonic suppression system 300 integrating multimodal data can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the converter harmonic suppression system 300 integrating multimodal data can also be one of many hardware modules of the wireless terminal.

[0049] Alternatively, in another example, the converter harmonic suppression system 300 that integrates multimodal data and the wireless terminal can also be separate devices, and the converter harmonic suppression system 300 that integrates multimodal data can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0050] Those skilled in the art will understand that the above embodiments are specific implementations of the present invention, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A converter harmonic suppression method integrating multimodal data, characterized in that, include: Acquire raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream; Feature engineering is performed on the original three-phase current data stream, original three-phase voltage data stream, original DC bus voltage data stream, original temperature data stream, and original load command asynchronous data stream to obtain the time series of state feature vectors; Time series risk prediction based on state feature vectors to obtain trigger signals; In response to the trigger signal being true, causal attribution and candidate action generation are performed on the current state feature vector to obtain a candidate action set; The candidate action set and the current state feature vector are input into the trained dynamic model to obtain the optimal avoidance action; The optimal avoidance action is analyzed to obtain new control parameters, and the converter is smoothly scheduled using the new control parameters.

2. The converter harmonic suppression method based on multimodal data according to claim 1, characterized in that, The state feature vector includes the RMS current, total harmonic distortion of the grid voltage, DC bus voltage ripple, load change rate, and estimated IGBT junction temperature.

3. The converter harmonic suppression method based on multimodal data according to claim 1, characterized in that, Temporal risk prediction based on time series data of state feature vectors to obtain trigger signals includes: The time series of state feature vectors is subjected to state sequence temporal feature encoding to obtain the final hidden state vector; Risk probability regression and trigger signal generation are performed on the final hidden state vector to obtain the trigger signal.

4. The converter harmonic suppression method based on multimodal data according to claim 3, characterized in that, Risk probability regression and trigger signal generation are performed on the final hidden state vector to obtain the trigger signal, including: The risk probability is obtained by performing nonlinear mapping and probability transformation on the final hidden state vector; Threshold decisions are made based on a comparison between the risk probability and a preset risk threshold to obtain a trigger signal.

5. The converter harmonic suppression method based on multimodal data according to claim 1, characterized in that, Causal attribution and candidate action generation are performed on the current state feature vector to obtain a candidate action set, including: Based on the causal graph, the main causal impact quantization is performed on the current state feature vector to obtain the main causal impact vector; Based on the causal graph, the cocausal impact quantization of the current state feature vector is performed to obtain the cocausal impact matrix. Dominant causal pattern recognition and candidate action generation are performed based on the main causal impact vector, cocausal impact matrix and action mapping table to obtain a candidate action set.

6. The converter harmonic suppression method based on multimodal data according to claim 5, characterized in that, Based on the causal graph, the main causal impulse quantization is performed on the current state feature vector to obtain the main causal impulse vector, including: The main causal impact vector is determined using the following formula: ； in, This represents the state feature vector at the current moment. State characteristic variables The current observation value, This indicates an event where harmonics exceed the standard. This indicates that a harmonic exceedance event has occurred. This represents the baseline or prior probability of a harmonic exceedance event. Indicates will The value is fixed as The probability of time harmonic exceedance events. Indicates that in a given hour Harmonic exceedance events The primary cause and effect.

7. The converter harmonic suppression method based on multimodal data according to claim 5, characterized in that, Based on the causal graph, the cocausal impact quantization of the current state feature vector is performed to obtain the cocausal impact matrix, including: The co-causal impact matrix is ​​determined using the following formula: ； ； ； in, This represents the state feature vector at the current moment. State characteristic variables The current observation value, This represents the state feature vector at the current moment. State characteristic variables The current observation value, This indicates an event where harmonics exceed the standard. This represents the baseline or prior probability of a harmonic exceedance event. Indicates will The value is fixed as At the same time The value is fixed as The probability of time harmonic exceedance events. express The corresponding primary causal impact, express The corresponding primary causal impact, express and Harmonic exceedance events The original synergistic effect, They represent , The value after Z-score standardization Indicates the scaling factor. Indicates the intervention condition is and State coupling weights at time, express and Co-causal impact.

8. The converter harmonic suppression method based on fused multimodal data according to claim 5, characterized in that, Dominant causal pattern recognition and candidate action generation are performed based on the main causal impact vector, cocausal impact matrix, and action mapping table to obtain a candidate action set, including: Based on the main causal impact vector and the cocausal impact matrix, the current dominant factors are determined; The candidate action set is generated based on the current dominant factors and action mapping table.

9. The converter harmonic suppression method based on fused multimodal data according to claim 8, characterized in that, Based on the main causal impact vector and the cocausal impact matrix, the current dominant factors are identified, including: The current dominant factor is determined using the following formula: ； in, This indicates the current dominant factor. The th element in the current state feature vector State characteristic variables The corresponding primary causal impact, The first in the co-causal impact matrix Line number The elements of the column are also and Co-causal impact, This represents the state feature vector at the current moment. There are 1 state feature variables, where D is the dimension of the state feature vector at the current moment.

10. A converter harmonic suppression system integrating multimodal data, characterized in that, include: The data acquisition module is used to acquire raw three-phase current data stream, raw three-phase voltage data stream, raw DC bus voltage data stream, raw temperature data stream, and raw load command asynchronous data stream; The feature engineering module is used to perform feature engineering on the original three-phase current data stream, the original three-phase voltage data stream, the original DC bus voltage data stream, the original temperature data stream, and the original load command asynchronous data stream to obtain the time series of state feature vectors; The temporal risk prediction module is used to predict temporal risks based on the time series of state feature vectors in order to obtain trigger signals. The causal attribution and candidate action generation module is used to perform causal attribution and candidate action generation on the current state feature vector in response to the trigger signal being true, so as to obtain a set of candidate actions. The optimal avoidance action dynamic generation module is used to input the candidate action set and the current state feature vector into the trained dynamic model to obtain the optimal avoidance action. The action analysis and smooth scheduling module is used to analyze the optimal avoidance action to obtain new control parameters, and then use the new control parameters to perform smooth scheduling of the converter.

Images