An adaptive modal injection type current source converter energy efficiency optimization control method based on a PPO algorithm

Abstract

The application provides a kind of adaptive modal injection type current source converter energy efficiency optimization control method based on PPO algorithm, it is related to power supply system field, it is applied to injection type current source converter, injection type current source converter includes twelve pulse inverter and multiple injection branches cascaded in the DC side of twelve pulse inverter, and the method comprises: collecting real-time meteorological data and real-time power grid state data;Through the reinforcement learning agent based on proximal policy optimization algorithm, based on real-time meteorological data and real-time power grid state data, determine the optimal action, wherein the optimal action at least includes target level number, the number of injection branches that need to be started and the number of injection branches that need to be disabled;According to optimal action, control multiple injection branches run, with the advantage of realizing the optimal operation of injection type current source converter under all operating conditions.

Description

Technical Field

[0001] This invention relates to the field of power supply systems, and in particular to an adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm. Background Technology

[0002] The penetration rate of new energy sources such as photovoltaic and wind power in the power grid is constantly increasing. However, the characteristics of new energy power generation differ significantly from those of traditional thermal power, exhibiting obvious intermittency and strong weather coupling. Photovoltaic power generation is highly susceptible to cloud movement, atmospheric aerosol concentration, and ambient temperature. Under cloudy weather conditions, the output power of a photovoltaic array may drop sharply within seconds, only to rebound instantly after the clouds move away. Wind power generation is highly random; gusts, wind shear, and wake effects can cause violent and disorderly fluctuations in output power, with power fluctuations ranging from 0% to 100%. This wide range of power fluctuations dominated by meteorological factors means that grid-connected converters are constantly operating under dynamically changing load conditions, rather than the ideal rated conditions designed in, posing a significant challenge to the grid connection of new energy sources.

[0003] Injection-type current source converters (RMCSCs) have become the preferred topology for high-power renewable energy grid connection and high-voltage direct current (HVDC) transmission due to their unique advantages. They eliminate the need for large-capacity electrolytic capacitors, possess DC-side fault-blocking capability, and achieve high-quality waveform output without grid-side filtering. To meet stringent grid harmonic standards (such as IEEE 519), mainstream RMCSCs typically employ multi-level injection technology. By connecting multiple auxiliary injection branches in parallel on the DC side, auxiliary switching devices are used to synthesize 7-level, 9-level, or even higher-step current waveforms, achieving excellent harmonic characteristics at extremely low switching frequencies.

[0004] Although injection-type current source converters are relatively mature, existing control strategies have significant shortcomings when facing weather-driven power fluctuations in renewable energy generation. On one hand, "fixed-mode" control limits further improvements in light-load efficiency. Injection-type current source converters utilize a fundamental-wave commutated main bridge in conjunction with a low-frequency auxiliary injection circuit, exhibiting low switching losses. However, traditional control strategies often employ a single, fixed highest-order mode, such as maintaining 7-level operation throughout the entire time. When weather conditions worsen and the system operates in the low-to-medium power range, although the DC current decreases and the absolute loss does not increase significantly, all auxiliary branch switching devices still need to continuously perform switching commutation to maintain a refined 7-level waveform. The switching frequency of the auxiliary branch switches is six times the power frequency. From an efficiency optimization perspective, the marginal benefit of waveform improvement from high-order injection under light-load conditions diminishes, while basic operating costs such as drive losses, switching losses, and high-frequency iron losses in magnetic components still exist. This unnecessary redundancy restricts the system from achieving maximum efficiency across the entire load range. On the other hand, fault-tolerant level switching and degradation operation technologies lack initiative. Some scholars have proposed this technology primarily to improve system reliability. By monitoring the hardware status of the injection circuit in real time, when an irreversible fault is detected in a branch, the control system triggers protection logic, disconnects the faulty branch, and reconstructs the pulse sequence of the remaining healthy branches, forcing the system to reconstruct into a low-order mode to continue operation. This sacrifices some waveform quality in exchange for uninterrupted operation under fault conditions. However, this technology only addresses fault states and cannot proactively adapt to power fluctuations for optimized control in a new energy grid environment.

[0005] Therefore, there is a need to provide an adaptive modal injection current source converter energy efficiency optimization control method based on the PPO algorithm to achieve optimal operation of the injection current source converter under all operating conditions. Summary of the Invention

[0006] This invention provides an adaptive modal injection current source converter energy efficiency optimization control method based on the PPO algorithm, applied to an injection current source converter, which includes a twelve-pulse converter and multiple injection branches cascaded on the DC side of the twelve-pulse converter. The method includes: collecting real-time meteorological data and real-time grid status data; determining the optimal action based on the real-time meteorological data and real-time grid status data using a reinforcement learning agent based on a near-end policy optimization algorithm, wherein the optimal action includes at least the target level number, the number of the injection branch to be activated, and the number of the injection branch to be disabled; and controlling the operation of multiple injection branches according to the optimal action.

[0007] Furthermore, the optimal action is determined by a reinforcement learning agent based on a near-end policy optimization algorithm, based on real-time meteorological data and real-time power grid status data. This includes: determining the target level based on the output power prediction value; and determining the number of injection branches to be activated and the number of injection branches to be disabled based on the target level, real-time meteorological data, and real-time power grid status data.

[0008] Further, determining the target level number based on the predicted output power includes: calculating a first mode switching threshold and a second mode switching threshold based on the average real-time junction temperature of the switching devices in the injection current source converter; determining a first level number based on the real-time load current, the first mode switching threshold, and the second mode switching threshold; determining a second level number based on the predicted output power; and determining the target level number based on the first level number and the second level number.

[0009] Furthermore, the reward function of the reinforcement learning agent based on the proximal policy optimization algorithm includes a basic performance reward function and a branch energy efficiency cost function.

[0010] Furthermore, the basic performance reward function is: , in, Basic performance bonus, This represents the total real-time loss. The rated power of the converter, This serves as the harmonic constraint reference value. For real-time network-side harmonic distortion rate, This refers to the DC bus voltage fluctuation rate. For meteorological confidence level, , , For weights.

[0011] Furthermore, the first weight is determined based on the DC-side load rate and the highest real-time junction temperature of the switching devices in the injection current source converter.

[0012] Furthermore, the second weight is determined based on the total harmonic distortion of voltage and the distortion of output current.

[0013] Furthermore, the branch energy efficiency cost function is: , in, For the sake of branch road energy efficiency, For the first The state of the injection branch's switching action. Let be the real-time energy efficiency cost of the k-th injection branch at time t. This represents the total number of injected branches.

[0014] Furthermore, the real-time energy efficiency cost of the injection branch at time t is determined based on the real-time loss estimate and the real-time junction temperature of the power device.

[0015] Furthermore, the twelve-pulse converter includes a Y-phase three-phase full-bridge circuit and a D-phase three-phase full-bridge circuit; The injection branch includes a Y-bridge injection switch connected to the Y three-phase full-bridge circuit, a D-bridge injection switch connected to the D three-phase full-bridge circuit, a first diode connected to the Y-bridge injection switch, a second diode connected to the D-bridge injection switch, and an injection inductor.

[0016] Compared with existing technologies, the adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm provided by this invention has at least the following beneficial effects: 1. In new energy power generation systems, converters mostly operate in low-to-medium load ranges. Existing technologies maintain a high-frequency, multi-branch highest-order injection mode throughout the entire lifecycle, resulting in low weighted average efficiency. This method, however, collects real-time meteorological and grid status data and utilizes a reinforcement learning agent based on a near-end policy optimization algorithm to determine the optimal action, encompassing the target level and the start / stop numbers of injection branches. It can dynamically adjust according to load conditions, reducing unnecessary switching losses, avoiding high-frequency switching losses in injection branches, minimizing energy waste, improving the system's weighted average efficiency across the entire load domain, and effectively solving the problem of overall energy efficiency collapse.

[0017] 2. The existing control system is binary, encountering difficulties in the medium load transition zone. Switching to the basic mode leads to deterioration of grid-side current harmonics, causing grid connection safety risks; maintaining the full-level mode results in excessive performance and low energy efficiency. This method utilizes a reinforcement learning agent to determine the optimal action based on real-time data. It can flexibly adjust the target level and injection branch operating status, finding suitable level and injection branch combinations under different load conditions. This ensures harmonic compliance while avoiding excessive performance, achieving a dynamic optimal balance between energy efficiency and power quality across the entire load domain.

[0018] 3. Under traditional control, the auxiliary switching devices in the injection circuit operate at high frequency throughout the entire cycle, accelerating the aging of power devices, leading to high-frequency iron losses in the injection transformer / inductor core, shortening the lifespan of auxiliary branches, reducing the mean time between failures (MTBF) of the entire system, and increasing maintenance costs. This method determines the optimal operation of the injection branches, reasonably disables some injection branches, reduces high-frequency operation of outer devices, reduces thermal shock and voltage stress, extends the lifespan of auxiliary devices, improves the reliability of the entire system, and reduces long-term maintenance costs. Attached Figure Description

[0019] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is a circuit diagram of an injection current source converter shown in some embodiments of this specification; Figure 2 This is a waveform diagram of the DC-side injected current of an injection-type current source converter according to some embodiments of this specification; Figure 3 This is a schematic diagram of a 6-level injection pulse unit according to some embodiments of this specification; Figure 4 This is a schematic diagram illustrating the structure of a full-cycle cyclic pulse according to some embodiments of this specification; Figure 5 This is a flowchart illustrating an adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm, according to some embodiments of this specification. Figure 6 This is a schematic diagram illustrating the three-level to five-level switching according to some embodiments of this specification; Figure 7 This is a schematic diagram illustrating the five-level to seven-level switching according to some embodiments of this specification; Figure 8 This is a schematic diagram of total harmonic distortion at three levels as shown in some embodiments of this specification; Figure 9 This is a schematic diagram of total harmonic distortion at five levels as shown in some embodiments of this specification; Figure 10 This is a schematic diagram of total harmonic distortion at seven levels as shown in some embodiments of this specification. Detailed Implementation

[0020] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0021] An adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm can be applied to... Figure 1The injection-type current source converter shown includes a twelve-pulse converter and multiple injection branches cascaded on the DC side of the twelve-pulse converter. The twelve-pulse converter is connected to the upstream transformer, which is electrically connected to the output terminal of photovoltaic, wind power, and other new energy power generation equipment. The three-phase bridge connected to the Y-type winding of the upstream transformer is defined as a Y-phase three-phase full-bridge circuit, and the three-phase bridge connected to the D-type winding of the upstream transformer is defined as a D-phase three-phase full-bridge circuit. The injection branches include the Y-bridge injection switches (i.e.,...) connected to the Y-phase three-phase full-bridge circuit. Figure 1 middle One of them), the D-bridge injection switch connected to the D three-phase full-bridge circuit (i.e. Figure 1 middle One of them), injected inductor (i.e. Figure 1 middle One of them) and diode (i.e. Figure 1 middle One of them). The turns ratio of the transformer grid-side winding to the valve-side winding is (Y-type winding) and (D-type winding). The voltage vector of the Y-type winding may lead or lag the voltage vector of the D-type winding. Switches in a three-phase full-bridge circuit and The device used is a semi-controlled thyristor, and the switch of the injection branch is connected to the DC bus. and It is composed of IGBTs, which are reverse-resistance devices. For the diode connected in the injection branch, and These are the A-phase currents of the Y-bridge and D-bridge connected to the transformer valve side, respectively. This refers to the A-phase current on the transformer grid side. and These are the injection currents of the Y-bridge and D-bridge, respectively. It is the DC side current. It is the output DC voltage. It is injected inductance. and The output DC voltage of the Y-bridge and D-bridge is different.

[0022] Figure 2 This is a schematic diagram of the waveform of the DC-side injected current of an injection-type current source converter according to some embodiments shown in this specification, such as... Figure 2As shown, the core principle of injection technology is based on the theories of "DC ripple injection" and "waveform synthesis." Essentially, it introduces a DC distribution unit (injection branch) consisting of auxiliary switches and energy storage elements onto the DC side of a traditional twelve-pulse converter (composed of two six-pulse bridges with a 30° phase difference). This unit operates synchronously at six times the grid fundamental frequency (e.g., 300Hz), actively "cutting" and reconstructing the originally flat DC current into a multi-level stepped wave (approximately a triangular wave) with a specific pattern. When this shaped DC stepped wave flows through the main bridge switching devices (such as thyristors), it utilizes waveform superposition and phase synchronization... The principle of cancellation multiplies the current steppedness induced on the grid side (e.g., from 12 pulses to 72 pulses), thereby automatically canceling the low-order characteristic harmonics (such as the 11th and 13th) on the AC side, synthesizing a high-quality approximate sine wave that meets grid connection standards without the need for a large-capacity filter. At the same time, the injected waveform is precisely designed to include a periodic "zero value interval", which creates a forced artificial zero current (ZCS) commutation condition for the high-power semi-controlled devices of the main bridge, enabling them to achieve lossless self-turn-off without relying on the natural zero crossing of the grid voltage, completely solving the commutation failure problem and significantly improving system efficiency.

[0023] Figure 3 This is a schematic diagram of a 6-level injection pulse unit according to some embodiments of this specification, such as... Figure 3 As shown, taking a six-level full-cycle cyclic pulse scheme as an example, according to the harmonic injection theorem, it is known that the injected current waveform and the corresponding superimposed shape of the injection switch pulse are consistent. Within one injection cycle (60°), with 30° as the axis of symmetry, the six-step trigger pulse can be decomposed into 10 different pulse units. The preceding... The five pulse units with the same height but different lengths are named , the last 30 ° The five injection pulse units are named Each pulse unit corresponds one-to-one with the pulse units decomposed in the first half. A cyclic pulse refers to a trigger pulse scheme for an injection circuit, which is defined as follows: for five injection switches connected to the same main bridge, the trigger pulse of any one group of injection switches is taken as the reference, and the pulses of the other four injection switches can be obtained by phase shifting them, as shown in Table 1.

[0024]

[0025] Figure 4 This is a schematic diagram illustrating the structure of a full-cycle cyclic pulse according to some embodiments of this specification, such as... Figure 4 As shown, the black portion represents the five IGBTs connected to the Y-bridge. The trigger pulses are complementary between the D-bridge and Y-bridge trigger pulses; the red pulse segment represents the five IGBTs connected to the D-bridge. The trigger pulse.

[0026] Figure 5 This is a flowchart illustrating an adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm, as shown in some embodiments of this specification. Figure 5 As shown, an adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm may include the following steps.

[0027] Step 510: Collect real-time meteorological data and real-time power grid status data.

[0028] Specifically, real-time meteorological data refers to data related to meteorological factors affecting the power generation capacity of new energy power generation equipment such as photovoltaics and wind power. Real-time meteorological data is a collection of various parameters reflecting the atmospheric conditions at the current moment, such as irradiance, cloud imagery, and wind speed. Among them, irradiance reflects the intensity of solar radiation, which is directly related to the amount of solar energy acquired. For solar energy utilization systems, it is a core indicator for assessing the scale of energy input. Cloud imagery presents the distribution, shape, and dynamic changes of clouds in the sky. By analyzing cloud imagery, changes in illumination conditions can be predicted in advance, providing a forward-looking basis for adjusting system operation strategies. Wind speed data is crucial for wind-related systems. The magnitude, direction, and stability of wind speed determine the operating status and power generation efficiency of wind power generation equipment. Real-time monitoring of wind speed helps ensure stable output of wind power generation.

[0029] Real-time grid status data is a collection of data used to describe various performance indicators of the power grid during real-time operation. For example, grid-side voltage / frequency deviation (ΔV, Δf) is a key parameter for measuring the stability of grid voltage and frequency. Abnormal fluctuations in voltage and frequency may affect the normal operation of electrical equipment and even threaten the safety of the entire power grid. Real-time monitoring of these two parameters can promptly detect abnormal conditions in the power grid. System real-time efficiency (η) directly reflects the system's ability to convert input energy into useful output during operation. By collecting its data in real time, we can understand the current operating efficiency of the system so that timely optimization and adjustments can be made. Output current harmonic distortion rate (Total Harmonic Distortion, THD) reflects the content of harmonic components in the output current. Excessive harmonic distortion rate will reduce power quality and have adverse effects on other equipment in the power grid. Real-time monitoring of this indicator helps to ensure the quality of output power.

[0030] Step 520: The optimal action is determined by a reinforcement learning agent based on a near-end policy optimization algorithm, using real-time meteorological data and real-time power grid status data.

[0031] The optimal action includes at least the target level number, the number of the injection branch to be activated, and the number of the injection branch to be disabled.

[0032] Specifically, the state space is a collection of environmental information perceived by the agent, consisting of real-time meteorological data and real-time power grid state data. The action space is the set of all actions that the agent can take, and each action must at least include the target level number, the number of injection branches to be activated, and the number of injection branches to be disabled. The target level number can be 3 levels, 5 levels, or 7 levels, etc.; the activation and deactivation of injection branches can flexibly adjust the current injection path in the power grid, determine the number of injection branches in operation, and thus change the step density of the injected current.

[0033] In some embodiments, step 520 specifically includes: The reinforcement learning agent based on the near-end policy optimization algorithm determines the target level number based on the output power prediction value; The reinforcement learning agent based on the near-end policy optimization algorithm determines the numbers of the injection branches that need to be activated and the numbers of the injection branches that need to be disabled based on the target level, real-time meteorological data, and real-time power grid status data.

[0034] Specifically, the output power prediction value can be the output power of new energy power generation equipment such as photovoltaic and wind power, predicted based on real-time meteorological data. The target power level can be determined by a reinforcement learning agent using a near-end policy optimization algorithm based on the output power prediction value in any manner. For example, the target power level can be determined based on preset rules. As an example only, the preset rules could include: when the output power prediction value is less than 30% of the rated power, the target power level can be 3 levels; when the output power prediction value is between 30% and 60% of the rated power, the target power level can be 5 levels; and when the output power prediction value is greater than 60% of the rated power, the target power level can be 7 levels.

[0035] In some embodiments, the reinforcement learning agent based on the proximal policy optimization algorithm determines the target level number based on the output power prediction value, including: The first mode switching threshold and the second mode switching threshold are calculated based on the average real-time junction temperature of the switching devices in the injection current source converter. The number of first levels is determined based on the real-time load current, the first mode switching threshold, and the second mode switching threshold. Determine the number of second levels based on the predicted output power value; The target number of levels is determined based on the first level number and the second level number.

[0036] Specifically, the first mode switching threshold and the second mode switching threshold can be calculated using the following formulas: , , in, The first mode switching threshold, The first reference switching current threshold can be, for example, 30% of the rated current. For thermal coupling weighting coefficients, These are the network-side coupling weight coefficients. The thermal protection drift factor characterizes the thermal stress margin of auxiliary branch components. , This represents the average real-time junction temperature of the switching devices in an injection current source converter. This refers to the rated reference operating temperature of the switching device (e.g., 80°C). This is the upper limit of the safe temperature for switching devices. When the average real-time junction temperature of the switching devices in the injection current source converter is detected to rise... Increasing this threshold drives a positive shift (upward movement) in the first and second mode switching thresholds. This means that even under a larger load current, the system will be forced to remain in a lower-order mode, delaying the activation of the high-frequency auxiliary branch and thus actively suppressing temperature rise. For power grid disturbance immunity drift factor, The harmonic distortion rate of the output current. This represents the maximum value of the output current harmonic distortion rate. As the output current harmonic distortion rate increases, Increasing the threshold value, since the pre-factor is negative, drives the first and second mode switching thresholds to drift negatively (downward). This means that the higher-order mode will be switched to earlier under smaller load currents, utilizing the higher-level waveforms to compensate for grid distortion. The threshold for the second mode switching. The second reference switching current threshold can be, for example, 60% of the rated current.

[0037] The first level number can be determined based on preset rules, real-time load current, first mode switching threshold, and second mode switching threshold.

[0038] For example, real-time load current When this occurs, it is determined that the system has entered the extremely light load energy-saving zone. At this time, the target voltage level can be 3 levels. When this occurs, it is determined that the system has entered the medium-load balanced zone. The target voltage level can be 5 levels. When this occurs, it is determined that the system has entered a heavy load or anti-interference operation zone. The target voltage level can be 7 levels, etc.

[0039] The number of second levels can be determined based on preset rules and the predicted output power value. As an example only, the preset rules may include: when the predicted output power value is less than 30% of the rated power, the number of second levels can be 3 levels; when the predicted output power value is between 30% and 60% of the rated power, the number of second levels can be 5 levels; when the predicted output power value is greater than 60% of the rated power, the number of second levels can be 7 levels.

[0040] The maximum value between the first level number and the second level number can be selected as the target level number.

[0041] Through the above process, the decision boundary is made flexible and adaptable to different operating conditions. When the device overheats, the power level is automatically reduced to protect the hardware. When the power grid is in poor condition, the power level is automatically increased to support grid connection. This significantly improves the system's environmental adaptability and robustness, ensuring that the converter can achieve a dynamic balance between energy efficiency and power quality under all operating conditions.

[0042] In some embodiments, the reward function of a reinforcement learning agent based on a proximal policy optimization algorithm includes a basic performance reward function and a branch energy efficiency cost function.

[0043] For example, the reward function can be: , in, For the total reward, Basic performance bonus, This is the proportionality constant, and its value is a positive number. This comes at the cost of improving the energy efficiency of branch roads.

[0044] The basic performance reward function mainly encompasses the positive reward for system efficiency and the penalty for exceeding the total harmonic distortion (THD) limit. Its core logic is that, when load conditions change, the guiding agent selects methods that maximize system efficiency while meeting grid-connected harmonic standards.

[0045] In some embodiments, the basic performance reward function is: , in, Basic performance bonus, This represents the total real-time loss, including both switching and conduction losses. The rated power of the converter, The harmonic constraint benchmark value is the upper limit threshold of total harmonic distortion (THD) specified in the grid connection standard (e.g., 5%). For real-time network-side harmonic distortion rate, This refers to the DC bus voltage fluctuation rate. That is, DC bus voltage With the rated value The difference is the same as the rated value. The ratio, with the negative sign before it, indicates a penalty for voltage fluctuations. This means that points are deducted for any instability, whether it's a voltage drop or a sudden voltage surge. The meteorological confidence level is used to suppress the algorithm's action range when the meteorological data is noisy. , , For weights.

[0046] In some embodiments, the first weight is determined based on the DC-side load rate and the highest real-time junction temperature of the switching devices in the injection current source converter.

[0047] For example, the first weight can be calculated using the following formula: , in, For example, the first weighted benchmark value. The value range can be [0.3, 0.6]. This is a scaling factor used to control the amplification factor of the first weight benchmark value, in order to prevent "gradient explosion" or "excitation failure" in reinforcement learning. The value should be a constant between 0.5 and 1.5, with a preferred value of 1.0. This is the real-time current on the DC side. This is the DC-side reference current. For the DC-side load factor, use the exponential decay function. Its characteristics enable it to prioritize reducing losses under light load conditions. This represents the highest real-time junction temperature of the switching devices in an injection current source converter. For ambient temperature, when the device temperature Approaching the limit hour, Approaching 0, leading to The decrease forces intelligent agents to stop blindly pursuing "high efficiency" when devices overheat, and instead tolerate a slight loss of efficiency in order to even out the heat, thus preventing the device from crashing.

[0048] In some embodiments, the second weight is determined based on the total harmonic distortion of voltage and the distortion of output current.

[0049] For example, the second weight can be calculated based on the following formula: , in, For example, the second weighted benchmark value. The value can range from [0.4, 0.8], with a preferred value of 0.5. This is the proportionality constant, with a value between 10 and 20. The total harmonic distortion of voltage. The distortion rate of the output current. The desired target distortion rate (e.g., 3%). The smoothing tolerance factor is the one for exponential penalty. Determined when real-time harmonics Approaching or exceeding the target value The degree of steepness at which the penalty weight increases. It acts as a "soft barrier" in the PPO algorithm, preventing control chattering caused by measurement noise. A constant with a value between 0.01 and 0.02.

[0050] when much smaller hour, Very small At lower levels, the agent can relax its quality requirements and focus on energy conservation. However, once... Approaching or exceeding , The exponential surge forced intelligent agents to sacrifice energy consumption for waveform quality.

[0051] This represents the severity of the penalty for system safety issues. It is usually set to a large positive value, meaning that any operation that causes a voltage drop will be severely punished.

[0052] , in, The conduction loss is proportional to the current amplitude. Switching losses are related to the target voltage level. The more target voltage levels there are, the more switching transistors are used, resulting in higher switching losses.

[0053] , in, This is a characteristic value of environmental volatility, representing the standard deviation or root mean square value of meteorological data within a past time window (such as the most recent 2 seconds, 5 minutes, 30 minutes, etc.). The larger the value, the more volatile the sunshine / wind speed, and the more unstable the weather. This serves as the baseline value for fluctuations, used for normalization adjustment. It is set based on local historical meteorological data, determining the algorithm's "sensitivity" to weather fluctuations. When the weather is stable... , The highest confidence level encourages the PPO to trust the prediction and take action. However, when the weather is extremely unstable (e.g., storms causing high sensor noise) and the prediction confidence is low, the prediction becomes less reliable. Approaching 0 or a negative value suppresses unnecessary frequent operations by PPO, keeping the system in its current state.

[0054] In some embodiments, the branch energy efficiency cost function is: , in, For the sake of branch road energy efficiency, For the switching action state of the k-th injection branch, : This indicates that the agent decides to participate in the work via the k-th injection branch. : Indicates that the agent decides to cut off (or bypass) the k-th injection branch. This represents the real-time energy efficiency cost of the k-th injection branch. It's a dynamically changing weighted value used to measure the cost of using this injection branch at the current moment. A smaller value indicates a healthier injection branch with lower losses, making it more worthy of priority selection. This represents the total number of injected branches.

[0055] In some embodiments, the real-time energy efficiency cost of the injection branch at time t is determined based on the real-time loss estimate and the real-time junction temperature of the power device.

[0056] For example, the real-time energy efficiency cost of the injection branch at time t can be calculated based on the following formula: , in, This is the real-time loss estimate for the k-th injection branch, calculated in real-time based on the current i flowing through this injection branch and the switching frequency f. It includes the device's conduction loss and switching loss. This refers to the real-time junction temperature sampled or estimated value of the power device (IGBT / IGCT) in the k-th injection branch. This is the rated reference temperature (e.g., 80°C) for the power device. This is the temperature normalization coefficient, used to adjust the sensitivity of the temperature penalty. and are the weighting coefficients for the loss term and the temperature rise term, respectively.

[0057] To meet the 5-level requirement, the agent must select a specific branch combination (e.g., select 2 branches from 3 redundant branches).

[0058] At this point, the intelligent system calculates the values ​​of all branches. value.

[0059] If the internal resistance of branch A increases due to aging ( (Increase), or due to poor heat dissipation leading to temperature rise. Increase, its The value will increase significantly.

[0060] according to The definition, in order to make the total reward Maximizing (i.e., minimizing penalty), intelligent agents tend to prioritize high... Branch state of value Set to 0, and lower The branch status of the value is set to 1.

[0061] In this way, the injection branches can operate in a non-rotating manner, instead employing adaptive energy-efficient optimal control.

[0062] Step 530: Control the operation of multiple injection branches according to the optimal action.

[0063] Specifically, for each level, the corresponding optimal switching angle table and the full-cycle cyclic pulse simplification matrix can be calculated and stored in advance.

[0064] The system can call the optimal switching angle table corresponding to the target level number and the full-cycle cyclic pulse simplification matrix to generate specific drive signals for the power devices in each injection branch that needs to be started. The grid voltage zero-crossing point is selected as the synchronization switching point. At the switching point, the system atomically executes: 1. Applying the new drive signal to the injection branch that needs to be started; 2. Blocking all drive signals for the injection branches that need to be disabled; 3. Adjusting the current controller reference value to smoothly decay the current in the injection branches that need to be disabled, while fine-tuning the current in the injection branches that need to be started to maintain the stability of the total DC current, achieving volt-second balance.

[0065] The following section, combined with experiments, illustrates the beneficial effects of an adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm. Taking a 7-level injection current source converter with a rated current of 1000A as an example, the first reference switching current threshold is set to 300A (30% of the rated current), and the second reference switching current threshold is set to 650A (60% of the rated current). The injection branches include injection branch 1, injection branch 2, and injection branch 3. At the beginning of operation, under a light load condition of 200A in the early morning, the reinforcement learning agent based on the near-end policy optimization algorithm determines that the system has entered the energy-saving zone, and instructs only injection branch 1 to operate in Mode I (3 levels), achieving extreme energy saving with zero auxiliary loss; as the load increases to 500A due to enhanced light, such as… Figure 6 As shown, the waveform smoothly evolves to Mode II (5 levels) and injection branches 1 and 2 are activated to achieve a balance between energy efficiency and waveform; by noon, under full load at 900A, as... Figure 7As shown, it automatically switches to Mode III (7-level) and activates injection branches 1, 2, and 3 to ensure high-quality grid connection with THD < 3%. Specifically, for sudden weather disturbances (such as afternoon cloud cover causing current drop but excessively high background harmonics, with measured THD > 5%), it features a closed-loop forced correction function. This function automatically triggers the switching to Mode III (7-level) logic, ignoring power saving commands and forcibly locking the higher-order mode to prioritize ensuring power quality meets national grid connection standards. Figure 8 , 9 As shown in Figure 10, the experimental results indicate that the THD value of the grid-side current decreases significantly with the increase of the number of injection levels: in the stable states of 3, 5, and 7 levels, the THD values ​​are 7.73%, 3.98%, and 2.77%, respectively. This represents a significant improvement over the traditional 12-pulse system. At this point, the grid-side current waveform is highly close to a sine wave, and low-order characteristic harmonics are effectively suppressed and shifted towards higher frequency bands such as the 71st and 73rd harmonics. This demonstrates that by adaptively selecting the number of injection levels, the energy utilization efficiency of the device can be improved while ensuring ease of engineering implementation, and the harmonic suppression performance of the system can be enhanced.

[0066] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. An adaptive modal injection current source converter energy efficiency optimization control method based on the PPO algorithm, characterized in that, The method, applied to an injection-type current source converter, wherein the injection-type current source converter includes a twelve-pulse converter and multiple injection branches cascaded on the DC side of the twelve-pulse converter, comprises: Collect real-time meteorological data and real-time power grid status data; The optimal action is determined by a reinforcement learning agent based on a near-end policy optimization algorithm, based on real-time meteorological data and real-time power grid status data. The optimal action includes at least the target level number, the number of the injection branch to be activated, and the number of the injection branch to be disabled. Based on the optimal action, control the operation of multiple injection branches.

2. The adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm according to claim 1, characterized in that, Using a reinforcement learning agent based on a near-end policy optimization algorithm, and based on real-time meteorological data and real-time power grid state data, the optimal action is determined, including: Determine the target level number based on the predicted output power value; The reinforcement learning agent based on the near-end policy optimization algorithm determines the numbers of the injection branches that need to be activated and the numbers of the injection branches that need to be disabled based on the target level, real-time meteorological data, and real-time power grid status data.

3. The adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm according to claim 2, characterized in that, The step of determining the target level number based on the predicted output power value includes: The first mode switching threshold and the second mode switching threshold are calculated based on the average real-time junction temperature of the switching devices in the injection current source converter. The number of first levels is determined based on the real-time load current, the first mode switching threshold, and the second mode switching threshold. Determine the number of second levels based on the predicted output power value; The target number of levels is determined based on the first level number and the second level number.

4. The adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm according to claim 2, characterized in that, The reward function of the reinforcement learning agent based on the proximal policy optimization algorithm includes a basic performance reward function and a branch energy efficiency cost function.

5. The adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm according to claim 4, characterized in that, The basic performance reward function is: ， in, Basic performance bonus, This represents the total real-time loss. The rated power of the converter, This serves as the harmonic constraint reference value. For real-time network-side harmonic distortion rate, This refers to the DC bus voltage fluctuation rate. For meteorological confidence level, , , For weights.

6. The adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm according to claim 5, characterized in that, The first weight is determined based on the DC-side load rate and the highest real-time junction temperature of the switching devices in the injection current source converter.

7. The adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm according to claim 5, characterized in that, The second weight is determined based on the total harmonic distortion of voltage and the distortion of output current.

8. The adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm according to any one of claims 4-7, characterized in that, The energy efficiency cost function of the branch is: ， in, For the sake of branch road energy efficiency, For the first The state of the injection branch's switching action. Let be the real-time energy efficiency cost of the k-th injection branch at time t. This represents the total number of injected branches.

9. The adaptive mode injection current source converter energy efficiency optimization control method based on the PPO algorithm according to claim 7, characterized in that, The real-time energy efficiency cost of the injection branch at time t is determined based on the real-time loss estimate and the real-time junction temperature of the power device.

10. A method for energy efficiency optimization control of an adaptive modal injection current source converter based on the PPO algorithm according to any one of claims 1-3, characterized in that, The twelve-pulse converter includes a Y-phase three-phase full-bridge circuit and a D-phase three-phase full-bridge circuit. The injection branch includes a Y-bridge injection switch connected to the Y three-phase full-bridge circuit, a D-bridge injection switch connected to the D three-phase full-bridge circuit, a first diode connected to the Y-bridge injection switch, a second diode connected to the D-bridge injection switch, and an injection inductor.

Images