Dynamic cooperative control system and method for gas turbine and microgrid

Abstract

The invention belongs to the field of data processing, and particularly relates to a dynamic cooperative control system and method for a gas turbine and a micro-grid, and the method comprises the steps: constructing a micro-grid real-time monitoring module, continuously collecting distributed energy real-time output, controllable load demands, bus voltage frequency and equipment state parameters, and carrying out the filtering and noise reduction through a preprocessing unit, thereby guaranteeing the data precision; calculating a real-time power difference value based on the preprocessed data, calling an adaptive neural fuzzy inference system, taking the power difference value, the bus voltage deviation and the frequency deviation as input, and judging whether the power difference value, the bus voltage deviation and the frequency deviation exceed a preset threshold value by means of a fuzzy rule base and a neural network model; if the threshold values are not exceeded, the current states of the gas turbine and the energy storage system are maintained; if any one exceeds the threshold value, a dynamic cooperative control instruction is triggered, precise cooperative control of the gas turbine and the micro-grid is achieved, and the operation stability, the operation efficiency and the reliability of the micro-grid are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of cooperative control, and particularly relates to a dynamic cooperative control system and method of a gas turbine and a microgrid. BACKGROUND

[0002] The gas turbine is a heat engine that converts the chemical energy of natural gas, biogas and the like into mechanical energy and then outputs electric energy, and is composed of a compressor, a combustion chamber and a turbine and follows the Brayton cycle. According to power, the gas turbine can be divided into a micro gas turbine adapted to a community / park microgrid and a small gas turbine adapted to an industrial park microgrid. Because of fast startup and high stability, the gas turbine is often used as a stable power supply in a microgrid to make up for the intermittency of renewable energy such as photovoltaic and wind power. The microgrid is a small autonomous power system containing distributed energy (gas turbine, photovoltaic, etc.), energy storage systems, controllable loads and monitoring and protection devices, and can realize grid-connected (bidirectional power exchange with a large power grid) or off-grid (island mode to guarantee critical loads) operation. The core goal is to improve power supply reliability, accommodate renewable energy and optimize energy utilization efficiency, and both need to maintain microgrid energy supply and demand balance through dynamic cooperative control.

[0003] In related technologies, from the core needs of microgrid operation, voltage stability and frequency stability are the basic premise for guaranteeing normal power supply of the microgrid, and the key to this premise is whether power fluctuations can be quickly responded to. In the current cooperative control scheme of the gas turbine and the microgrid, the response speed of the gas turbine is insufficient. Although the gas turbine starts faster than a traditional thermal power unit, it is still difficult to respond to millisecond-level and second-level power fluctuations such as sudden output reduction of photovoltaic / wind power and sudden increase of load, and the limited output adjustment range of the micro gas turbine further limits its rapid energy compensation capability, and it cannot quickly suppress short-term severe fluctuations. Therefore, there is an urgent need for a dynamic cooperative control scheme of the gas turbine and the microgrid to solve the above problems, so as to improve the operation stability, efficiency and reliability of the microgrid. SUMMARY

[0004] The application provides a dynamic cooperative control system and method of a gas turbine and a microgrid, which can improve the operation stability, efficiency and reliability of the microgrid.

[0005] In a first aspect, the application provides a dynamic cooperative control method of a gas turbine and a microgrid, comprising: A microgrid real-time monitoring module is constructed to continuously collect real-time output of distributed energy in the microgrid, real-time demand of controllable load, microgrid bus voltage and frequency, gas turbine operating state parameters and energy storage system charging and discharging state parameters. The distributed energy at least includes photovoltaic power, wind power and a gas turbine, and the gas turbine at least includes a micro gas turbine adapted to a community / park microgrid or a small gas turbine adapted to an industrial park microgrid. The real-time monitoring module is equipped with a data preprocessing unit to denoise high-frequency fluctuation data collected by a filtering algorithm to ensure that the data accuracy meets the subsequent control requirements. The real-time power difference of the microgrid is calculated based on the preprocessed data, and the real-time power difference is a difference between total output of the distributed energy and total demand of the controllable load. A preset adaptive neuro-fuzzy inference system (ANFIS) is called to take the real-time power difference, the bus voltage deviation value of the microgrid, and the frequency deviation value as input variables, to determine whether the real-time power difference exceeds a preset power fluctuation threshold value through a fuzzy rule base and a neural network training model of the ANFIS, and to determine whether the bus voltage deviation value of the microgrid or the frequency deviation value exceeds a preset stability threshold value. If none of the real-time power difference, the bus voltage deviation value of the microgrid, and the frequency deviation value exceeds, the current output of the gas turbine and the current charging and discharging state of the energy storage system are maintained; if any one of the real-time power difference, the bus voltage deviation value of the microgrid, and the frequency deviation value exceeds, a dynamic cooperative control instruction is triggered.

[0006] In a second aspect, an embodiment of the present application provides a dynamic cooperative control system of a gas turbine and a microgrid, which is applied to the dynamic cooperative control method of the gas turbine and the microgrid in the first aspect, and includes the following. A monitoring unit is configured to construct a real-time monitoring module of the microgrid, to continuously collect real-time output of distributed energy, real-time demand of controllable load, bus voltage and frequency of the microgrid, operating state parameters of the gas turbine, and charging and discharging state parameters of the energy storage system; the distributed energy at least includes photovoltaic power, wind power, and the gas turbine, and the gas turbine at least includes a micro gas turbine adapted to a community / microgrid or a small gas turbine adapted to an industrial park microgrid; the real-time monitoring module is loaded with a data preprocessing unit, to denoise high-frequency fluctuation data collected through a filtering algorithm, and to ensure that data accuracy meets subsequent control requirements; A control unit is configured to calculate a real-time power difference of the microgrid based on preprocessed data, and the real-time power difference is a difference between total output of the distributed energy and total demand of the controllable load; a preset adaptive neuro-fuzzy inference system (ANFIS) is called to take the real-time power difference, the bus voltage deviation value of the microgrid, and the frequency deviation value as input variables, to determine whether the real-time power difference exceeds a preset power fluctuation threshold value through a fuzzy rule base and a neural network training model of the ANFIS, and to determine whether the bus voltage deviation value of the microgrid or the frequency deviation value exceeds a preset stability threshold value; if none of the real-time power difference, the bus voltage deviation value of the microgrid, and the frequency deviation value exceeds, the current output of the gas turbine and the current charging and discharging state of the energy storage system are maintained; if any one of the real-time power difference, the bus voltage deviation value of the microgrid, and the frequency deviation value exceeds, a dynamic cooperative control instruction is triggered.

[0007] In a third aspect, an embodiment of the present application provides an electronic device, which includes: at least one processor, a memory, and an input-output unit; The memory is configured to store a computer program, and the processor is configured to invoke the computer program stored in the memory to execute the method for dynamic collaborative control of the gas turbine and the microgrid.

[0008] In a fourth aspect, a computer-readable storage medium is provided, which includes instructions that, when executed on a computer, cause the computer to perform the method for dynamic collaborative control of the gas turbine and the microgrid of the first aspect.

[0009] The technical scheme provided by the embodiments of the present application constructs a microgrid real-time monitoring module, continuously collects real-time output of distributed energy in the microgrid, real-time demand of controllable load, microgrid bus voltage and frequency, gas turbine operation state parameters, and energy storage system charging and discharging state parameters; the distributed energy at least includes photovoltaic power supply, wind power supply, and a gas turbine, and the gas turbine at least includes a micro gas turbine adapted to a community / park microgrid or a small gas turbine adapted to an industrial park microgrid; the real-time monitoring module is equipped with a data preprocessing unit, which reduces noise of collected high-frequency fluctuation data through a filtering algorithm, and ensures that data accuracy meets subsequent control requirements. Further, based on the preprocessed data, a microgrid real-time power difference value is calculated, which is a difference between total output of the distributed energy and total demand of the controllable load. A preset adaptive neuro-fuzzy inference system (ANFIS) is called, with the real-time power difference value, a microgrid bus voltage deviation value, and a frequency deviation value as input variables, and through a fuzzy rule base and a neural network training model of the ANFIS, it is determined whether the real-time power difference value exceeds a preset power fluctuation threshold value, and whether the microgrid bus voltage deviation value or the frequency deviation value exceeds a preset stability threshold value. If none of the real-time power difference value, the microgrid bus voltage deviation value, and the frequency deviation value exceeds, the current output of the gas turbine and the current charging and discharging state of the energy storage system are maintained; if any one of the real-time power difference value, the microgrid bus voltage deviation value, and the frequency deviation value exceeds, a dynamic collaborative control instruction is triggered.

[0010] The technical scheme of the present application, on the one hand, greatly improves the accuracy and timeliness of the microgrid operation state judgment through data preprocessing and ANFIS optimization, and solves the problems of data noise interference and one-sided judgment in traditional control. On the other hand, through a differentiated instruction triggering strategy, while ensuring the stability of the microgrid voltage and frequency and the balance of power supply and demand, the invalid adjustment loss of equipment is reduced, and the service life of the gas turbine and the energy storage system is prolonged. Ultimately, the microgrid operation stability, power supply reliability, and operation efficiency are collaboratively improved, providing key technical support for efficient collaboration of the gas turbine and the microgrid in different scenarios such as communities and industrial parks, especially adapting to the complex operation requirements of the microgrid after renewable energy such as photovoltaic and wind power is connected, and promoting the upgrade of the microgrid from passive stability maintenance to active precise control. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings: Figure 1 is a structural schematic diagram of a flow system of a dynamic cooperative control method of a gas turbine and a microgrid according to an embodiment of the application; Figure 2 is a structural schematic diagram of a dynamic cooperative control system of a gas turbine and a microgrid according to an embodiment of the application; Figure 3 is a structural schematic diagram of an electronic device according to an embodiment of the application. DETAILED DESCRIPTION

[0012] In order to make the objectives, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some but not all of the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0013] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0014] In view of at least one technical problem, the embodiments of the application provide a dynamic cooperative control system and method of a gas turbine and a microgrid.

[0015] The dynamic cooperative control scheme of the gas turbine and the microgrid provided by the embodiments of the application can also be executed by an electronic device. The electronic device can be a server, a server cluster, a cloud server. The electronic device can also be a terminal device such as a mobile phone, a computer, a tablet computer, a wearable device, or a special-purpose device (such as a special-purpose terminal device with a dynamic cooperative control system of a gas turbine and a microgrid, etc.). The above-mentioned chips introduced in the embodiments can also be carried in these electronic devices. Alternatively, these electronic devices can also install a service program for executing the dynamic cooperative control scheme of the gas turbine and the microgrid.

[0016] Figure 1 is a structural schematic diagram of a dynamic cooperative control method of a gas turbine and a microgrid according to an embodiment of the application, comprising the following steps: Step 101, a microgrid real-time monitoring module is constructed, and real-time output of distributed energy, real-time demand of controllable load, microgrid bus voltage and frequency, running state parameters of a combustion engine, and charging and discharging state parameters of an energy storage system are continuously collected; the distributed energy includes at least photovoltaic power, wind power and a combustion engine, and the combustion engine includes at least a micro combustion engine adapted to a community / park microgrid or a small combustion engine adapted to an industrial park microgrid; Step 102, the real-time monitoring module is equipped with a data preprocessing unit, high-frequency fluctuation data collected is denoised through a filtering algorithm, and data accuracy is ensured to meet subsequent control requirements; Step 103, a real-time power difference of the microgrid is calculated based on the preprocessed data, and the real-time power difference is a difference between total output of the distributed energy and total demand of the controllable load; Step 104, a preset adaptive neuro-fuzzy inference system (ANFIS) is called, real-time power difference, microgrid bus voltage deviation and frequency deviation are taken as input variables, a fuzzy rule base and a neural network training model of the ANFIS are used to determine whether the real-time power difference exceeds a preset power fluctuation threshold, and whether the microgrid bus voltage deviation or the frequency deviation exceeds a preset stability threshold; Step 105, if the real-time power difference, the microgrid bus voltage deviation and the frequency deviation do not exceed, the current output of the combustion engine and the current charging and discharging state of the energy storage system are maintained; if any of the real-time power difference, the microgrid bus voltage deviation and the frequency deviation exceeds, a dynamic cooperative control instruction is triggered.

[0017] In the embodiment of the application, in the dynamic cooperative control system of the combustion engine and the microgrid, the ANFIS is used as a core judgment unit, a fusion mechanism of fuzzy logic reasoning and neural network self-learning is used to realize accurate determination of the microgrid running state, and the processing process is closely combined with the microgrid parameters collected by the system.

[0018] Specifically, first, the input variable preprocessing and fuzzification. The ANFIS takes the real-time power difference, bus voltage deviation, and frequency deviation output by the real-time monitoring module in the system as the core input variables, normalizes the range of these parameters after filtering and noise reduction, and then converts them into fuzzy language variables (such as small power difference, medium voltage deviation, and large frequency deviation) through a pre-set membership function (such as Gaussian function), completing the conversion from accurate numerical value to fuzzy feature and adapting to the uncertainty of microgrid parameter fluctuations. Second, fuzzy rule base construction and matching. The rule base is generated based on historical operation data of the microgrid. Through fuzzy C-means clustering (FCM) algorithm, a large number of historical input variables and threshold judgment result samples are clustered to extract fuzzy rules that meet the working conditions of the microgrid, such as if the power difference exceeds the threshold and the voltage deviation exceeds the threshold, it is determined that the control needs to be triggered. In operation, the ANFIS will automatically match the corresponding rules in the rule base according to the fuzzy features of the current input variables to determine the fuzzy reasoning logic to which the current microgrid state belongs. Third, fuzzy reasoning calculation. The ANFIS performs reasoning by means of the forward propagation mechanism of neural networks. The fuzzy rule weight matched and the membership value of the input variable are multiplied, and then the reasoning results of all rules are integrated through fuzzy operators (such as product operator and max operator) to obtain the fuzzy judgment contribution value of each input variable, which comprehensively reflects the influence degree of the current power difference, voltage deviation, and frequency deviation on whether the threshold is exceeded. Finally, defuzzification and output of the judgment conclusion. The fuzzy contribution value obtained by reasoning is converted into an explicit numerical judgment result through defuzzification algorithms such as the center of gravity method. The binary conclusions of whether the real-time power difference exceeds the pre-set power fluctuation threshold, whether the bus voltage deviation or frequency deviation exceeds the pre-set stability threshold are output. If any of the conclusions is yes, the signal of triggering the dynamic cooperative control instruction is fed back to the system, and if all the conclusions are no, the signal of maintaining the current state of the gas turbine and energy storage is fed back, providing accurate basis for subsequent control decisions of the system.

[0019] Further optionally, the ANFIS will also continuously optimize the membership function parameters and rule weights through the back propagation algorithm to adapt to dynamic conditions such as photovoltaic / wind power output fluctuations and load changes in the microgrid, avoid the one-sidedness and false triggering problems of traditional fixed rule judgment, and improve the accuracy and robustness of the system in judging the operation state of the microgrid.

[0020] The specific implementation process of the present application will be described in detail below in conjunction with each step.

[0021] Step 101 is to construct a micro-grid real-time monitoring module and parameter collection. The principle is based on the full-dimensional state perception concept. Through the deployment of voltage sensors, current sensors, power transducers, gas turbine speed and temperature sensors, energy storage SOC sensors and other devices, the continuous collection of core operating parameters of the micro-grid is realized. Distributed energy output reflects energy supply capacity, controllable load demand reflects energy consumption demand, bus voltage and frequency are the core indicators of power quality in the micro-grid, and gas turbine and energy storage state parameters provide a basis for device adjustment. At the same time, the scene adaptability of distributed energy and gas turbine is clear, and the monitoring range covers all links of the micro-grid source, load, storage and network. In the community micro-grid, the real-time monitoring module can collect the real-time output of micro gas turbines, the output of photovoltaic arrays, the real-time demand of controllable loads such as air conditioners and charging piles, low-voltage bus voltage and frequency, and the charging and discharging current and voltage of lithium battery energy storage. In the industrial park micro-grid, the output of small gas turbines, wind farm output, production line motor load demand, high-voltage bus parameters and flywheel energy storage status can be collected. Thus, the problem of incomplete and wide coverage of traditional micro-grid monitoring parameters is solved, and the control decision is avoided due to the lack of key parameters. The full and real-time raw data support is provided for subsequent data processing and supply and demand balance judgment, and the micro-grid operation state is ensured to be perceived and known.

[0022] Step 102 is the filtering and noise reduction of the data preprocessing unit. The principle is that the original data collected by the micro-grid is easily disturbed by the environment, such as voltage fluctuation caused by electromagnetic interference, and high-frequency fluctuation of photovoltaic and wind power output caused by clouds or gusts. Through Kalman filtering, moving average filtering, wavelet filtering and other filtering algorithms, high-frequency noise is suppressed. The filtering algorithm separates the effective signal and the noise signal by constructing a data fluctuation model, such as the slow output change of photovoltaic, and the noise signal such as instantaneous electromagnetic interference. While retaining the trend characteristics of the data, meaningless fluctuations are eliminated to ensure that the data accuracy meets the threshold requirements of subsequent calculation and judgment. In the community micro-grid, the original data of photovoltaic array output may have high-frequency fluctuations for a short time due to cloud shading. After Kalman filtering, the fluctuation amplitude is significantly reduced, and the smoothed output data is more consistent with the actual energy supply trend. In the industrial park micro-grid, the start of the production line load may cause a short-term instantaneous fluctuation of the bus voltage. After wavelet filtering, the voltage data fluctuation range is reduced, avoiding noise interference with subsequent voltage deviation judgment. Thus, the data signal-to-noise ratio is improved, and the high-frequency noise is prevented from causing real-time power difference calculation deviation and voltage and frequency deviation misjudgment, such as avoiding noise from misjudging the normal power difference as exceeding the standard or misjudging the normal voltage as deviating from the threshold. Accurate and reliable preprocessed data is provided for subsequent steps, reducing the risk of false triggering of control decisions.

[0023] Step 103 is the micro-grid real-time power difference calculation. The principle is that the real-time power difference is the core quantitative index of the micro-grid supply-demand balance state. The difference between the total distributed energy output and the total controllable load demand is calculated. The total distributed energy output is the sum of photovoltaic, wind power and gas turbine output. The total controllable load demand is the sum of all controllable load power consumption. A positive difference indicates energy surplus, i.e. supply is greater than demand. A negative difference indicates energy deficiency, i.e. supply is less than demand. A difference close to 0 indicates that supply and demand are basically balanced. This index directly reflects whether the micro-grid is at risk of power imbalance. In a community micro-grid, the pre-processed data shows photovoltaic output, micro gas turbine output, total distributed energy output and total controllable load demand. The real-time power difference can be calculated to determine whether there is an energy surplus. In an industrial park micro-grid, the wind power output, small gas turbine output, total distributed energy output and production line load demand are calculated to obtain the real-time power difference to determine whether there is an energy deficiency. Thus, the abstract supply-demand relationship of the micro-grid is converted into a concrete numerical index, solving the subjective problem of traditional experience-based supply-demand judgment. The subsequent ANFIS threshold judgment has a clear quantitative basis, and provides a core decision reference for energy storage charging and discharging direction and gas turbine output adjustment. When there is an energy surplus, the energy storage is charged. When there is an energy deficiency, the energy storage is discharged and the gas turbine output is increased.

[0024] As an optional embodiment, in step 103, the micro-grid real-time power difference is calculated based on the pre-processed data. The real-time power difference is the difference between the total distributed energy output and the total controllable load demand, including: The distributed energy total output calculation model is constructed, the wavelet transform-Kalman filter fusion algorithm is introduced, the preprocessed distributed energy real-time output data output by the microgrid real-time monitoring module is subjected to secondary denoising and trend extraction; wherein, the wavelet transform unit separates the high-frequency noise and low-frequency effective signal in the output data through multi-scale decomposition, the Kalman filter unit dynamically filters the decomposed low-frequency effective signal, eliminates the fluctuations caused by data acquisition delay and environmental interference, and outputs the smooth output value of each distributed energy. The controllable load total demand calculation model is constructed, and the adaptive weight distribution algorithm is used to aggregate the preprocessed real-time demand data of each controllable load; according to the importance level and operating characteristics of each controllable load, the calculation weight of the demand data is dynamically allocated, the instantaneous value correction factor is additionally superimposed on the impact load demand data, the calculation deviation of the total demand caused by the sudden load demand is avoided, and the smooth value of the total demand of the controllable load is output. The power data time sequence alignment mechanism is established, through the time synchronization unit of the microgrid real-time monitoring module, the smooth output value of each distributed energy and the smooth value of the total demand of the controllable load are time stamp calibrated, it is ensured that all the power data participating in the calculation correspond to the same time dimension, and the calculation error caused by the asynchronous data acquisition is eliminated. The microgrid real-time power difference value is calculated, the smooth output value of each distributed energy is summed to obtain the total output of the distributed energy, and then the smooth value of the total demand of the controllable load is subtracted to obtain the real-time power difference value. The power difference dynamic correction mechanism is introduced, the power difference correction model is constructed based on the historical operation data of the microgrid, the current microgrid operation mode and the distributed energy output fluctuation coefficient are taken as the input variables, the radial basis function neural network RBFNN is used to output the power difference correction coefficient, the power difference correction coefficient is multiplied by the real-time power difference value to obtain the microgrid real-time power difference value. The microgrid real-time power difference value can reflect the actual energy supply and demand imbalance state of the microgrid, and provide reliable data support for subsequent ANFIS judgment and cooperative control instruction triggering.

[0025] In the above steps, when calculating the real-time power difference value of the microgrid based on the preprocessed data, a distributed energy total output calculation model is first constructed. The wavelet transform-Kalman filter fusion algorithm introduced in the model belongs to a multi-algorithm collaborative noise reduction technology. The wavelet transform separates high-frequency noise from low-frequency effective signals through multi-scale decomposition. The Kalman filter dynamically optimizes the low-frequency signals based on state estimation theory. The combination of the two can further improve the smoothness of the data. In the community microgrid scenario, the preprocessed distributed energy data of the microgrid monitoring module includes the output data of a 20kW micro gas turbine and a 50kW photovoltaic array. After processing by the wavelet transform-Kalman filter fusion algorithm, the high-frequency fluctuations in the photovoltaic output data caused by rapid cloud cover are further suppressed, and the instantaneous deviations in the micro gas turbine output data caused by slight fluctuations in fuel supply are eliminated. The final output is a smooth output value of 18kW for the micro gas turbine and 22kW for the photovoltaic array. In the industrial park microgrid scenario, the preprocessed output data of a 5MW small gas turbine and a 2MW wind farm are processed by the algorithm. The short-term fluctuations in wind power output caused by gusts are filtered, and the interference signals in the small gas turbine output caused by unit vibration are eliminated. The output is a smooth output value of 4.8MW for the small gas turbine and 1.9MW for the wind power.

[0026] Then a controllable load total demand calculation model is constructed. The adaptive weight distribution algorithm used dynamically adjusts the weight according to the importance and operating characteristics of the load, which belongs to intelligent load aggregation technology. The instantaneous value correction factor of the impact load is used to compensate for the impact of sudden loads on total demand calculation. In the community microgrid, controllable loads include residential air conditioners, charging piles, and lighting loads. Charging piles belong to impact loads. According to the importance level and operating characteristics, charging pile demand data is assigned a higher calculation weight, and an instantaneous value correction factor is added to avoid sudden changes in demand data caused by the simultaneous start of multiple charging piles at the same time. After processing by the model, the air conditioner load demand is 15kW, the charging pile demand is 8kW, and the lighting load demand is 3kW, with a smooth total demand value of 25kW for controllable loads. In the industrial park microgrid, controllable loads include production line motors, office loads, and auxiliary equipment loads. Production line motors belong to impact loads and are the most important. They are assigned the highest calculation weight and a correction factor. The final output is a production line motor demand of 4.5MW, an office load demand of 0.3MW, an auxiliary equipment load demand of 0.2MW, and a smooth total demand value of 5MW for controllable loads.

[0027] Then, the community microgrid sums the smooth output values of microturbine and photovoltaic to obtain the total distributed energy output 40 kW, and subtracts the total demand smooth value 25 kW of controllable load to obtain the real-time power difference basis value 15 kW. The industrial park microgrid sums the smooth output values of small turbine and wind power to obtain the total distributed energy output 6.7 MW, and subtracts the total demand smooth value 5 MW of controllable load to obtain the real-time power difference basis value 1.7 MW.

[0028] Finally, the power difference dynamic correction mechanism is introduced. The correction model based on the radial basis function neural network belongs to the nonlinear function approximation technology, and can dynamically adjust the correction coefficient according to the microgrid operation state. The community microgrid is currently in grid-connected mode, and the fluctuation coefficient of distributed energy output is low. The radial basis function neural network outputs a correction coefficient of 0.95, which is multiplied by the real-time power difference basis value 15 kW to obtain the final microgrid real-time power difference 14.25 kW. The industrial park microgrid is currently in off-grid mode, and the fluctuation coefficient of distributed energy output is high. The neural network outputs a correction coefficient of 1.02, which is multiplied by the basis value 1.7 MW to obtain the final real-time power difference 1.734 MW.

[0029] The microgrid real-time power difference calculated by the above process can more accurately reflect the actual energy supply and demand imbalance state of the microgrid. In the community microgrid, the supply and demand judgment deviation caused by photovoltaic fluctuation and charging pile impact is avoided, so that the ANFIS can judge the threshold based on the true supply and demand state. In the industrial park microgrid, the disturbance of wind power fluctuation and production line load mutation to the supply and demand judgment is eliminated, providing reliable data support for the ANFIS threshold judgment, thereby ensuring the accuracy of the dynamic collaborative control instruction trigger, avoiding control decision errors caused by power difference calculation deviation, and providing accurate basis for subsequent energy storage charging and discharging direction and turbine output adjustment, so that the adjustment action of energy storage and turbine is more in line with the actual demand of the microgrid.

[0030] As an optional embodiment, in step 103, the calculation weight of demand data is dynamically allocated according to the importance level and operation characteristics of each controllable load, and an instantaneous value correction factor is additionally added to the impact load demand data to avoid calculation deviation of total demand caused by single load burst demand, and the total demand smooth value of controllable load is output, including: A controllable load classification and feature extraction model is constructed, the VMD algorithm is used to decompose the real-time demand data of each controllable load after pretreatment, and the stationary component and the impact component in the load demand are separated; combined with the preset load importance level of the microgrid load management system, the importance coefficient and the impact characteristic coefficient of each load are extracted to form a load feature matrix; a dynamic weight allocation model is constructed based on DQN, the load feature matrix, the current operation mode of the microgrid, and the real-time output stability of distributed energy are used as state inputs, and the reward function is to minimize the total demand calculation deviation of controllable load, through the interaction training of DQN and microgrid operation environment, the mapping relationship between load characteristics and calculation weight is established; in the training process, the experience replay mechanism is used to optimize the weight update strategy, to ensure that the adaptive load demand calculation weight can be output under different working conditions; for the impact load demand data, an instantaneous value correction model is constructed; for the impact component obtained by VMD decomposition, the peak time, duration and attenuation characteristics of the impact load are extracted by HHT; based on the historical impact load response data of the microgrid, a correction factor library of impact load instantaneous value is established, the peak intensity and duration of the impact component are used as indexes to match the corresponding correction factor, the correction factor changes dynamically with the impact load attenuation trend, which avoids the over-correction or insufficient correction caused by a single fixed correction value; the weighted demand value of each controllable load is calculated, the calculation weight is multiplied by the real-time demand data of the corresponding load, and the instantaneous value correction factor is additionally superimposed on the impact load to obtain the weighted demand value of each controllable load; a controllable load total demand smoothing model is constructed, a sliding window filtering algorithm is used to aggregate and smooth the weighted demand value of each controllable load, and the window length is dynamically adjusted according to the load fluctuation frequency of the microgrid; a load demand trend prediction mechanism is introduced, the future load demand trend in a short time is predicted by GRU, and the trend of the aggregated total demand value is corrected to output the controllable load total demand smoothing value.

[0031] Specifically, in the construction of controllable load classification and feature extraction model, the core calculation process starts from the processing of the real-time demand data of each controllable load after pretreatment by VMD algorithm. Through iterative search for the optimal center frequency and bandwidth of each mode, the load demand data is decomposed into multiple independent modal components, from which the stationary component representing regular consumption and the impact component representing sudden fluctuation are separated. Then, combined with the preset load importance level of the microgrid load management system, the importance of different loads is converted into importance coefficients in specific numerical form according to the grade quantization rule. At the same time, the impact characteristic coefficient is calculated according to the proportion of the impact component in the total load demand, and the two coefficients are integrated according to the load type to form a load feature matrix with matching dimensions and load quantity. Based on the construction of dynamic weight allocation model by DQN, the core calculation takes the load feature matrix, the current operating mode parameters of the microgrid, and the real-time output stability indicators of distributed energy as the state space input. Through the neural network structure of DQN, the action value function under different states is calculated, and the reward function is minimized by the total demand calculation deviation of controllable load. In the interaction process with the microgrid operating environment, the network parameters are constantly updated. During training, the historical state action reward data is stored in the experience pool through the experience replay mechanism, and the data is randomly sampled for batch training. The weight update strategy is optimized to enable the model to output adaptive load demand calculation weights for different working conditions.

[0032] The core calculation of the instantaneous value correction model for impact load demand data starts from the processing of the impact component obtained by VMD decomposition through HHT. The impact component is decomposed into intrinsic mode functions through empirical mode decomposition, and then Hilbert transform is performed on each intrinsic mode function to obtain the instantaneous frequency and instantaneous amplitude of the impact load. Then, key parameters such as peak duration and decay characteristics are extracted. Based on the historical impact load response data of the microgrid, the influence of impact load on total demand calculation under different peak intensity and duration is statistically analyzed, and a library of impact load instantaneous value correction factors is established. During calculation, the peak intensity and duration of the current impact component are used as retrieval conditions to match the corresponding correction factor in the factor library, and the correction factor value is dynamically adjusted according to the change of instantaneous amplitude during the decay process of the impact load.

[0033] When calculating the weighted demand value of each controllable load, the calculated weight output by the dynamic weight distribution model is multiplied element by element with the real-time demand data of the corresponding load to obtain a basic weighted value. For the impact load, a transient value correction factor is superimposed on the basis of the weighted value, and the corrected value is integrated as the final weighted demand value of the impact load through addition operation. When constructing the total demand smoothing model of controllable load, the core calculation of the sliding window filtering algorithm is to set the initial window length, and then perform arithmetic average operation on the weighted demand value of each controllable load in the window to obtain a preliminary aggregated value. At the same time, the window length is adjusted according to the real-time change of the microgrid load fluctuation frequency. When the fluctuation frequency is high, the window length is shortened to increase the calculation frequency. When the fluctuation frequency is low, the window length is lengthened to reduce the consumption of calculation resources. When the load demand trend prediction mechanism is introduced, the historical load demand data is learned for time sequence characteristics through the gated recurrent unit GRU, and the predicted value of the load demand in the future short time is calculated. The predicted value is compared with the aggregated total demand value obtained by the sliding window filtering, and the trend correction amount is calculated according to the deviation degree. The aggregated value is corrected to obtain the final total demand smoothing value of controllable load.

[0034] After step 103, step 104 is the threshold judgment of ANFIS, which is similar to the foregoing description. The principle is that the adaptive neuro-fuzzy inference system ANFIS combines the uncertainty processing capability of fuzzy logic and the self-learning capability of neural network. The workflow includes four steps. The first step is to fuzz the input variables, and convert the real-time power difference, voltage deviation and frequency deviation into fuzzy language variables, such as large difference, medium difference and small difference. The second step is fuzzy rule matching, and a rule base is generated based on the historical data of the microgrid. For example, if the power difference exceeds the threshold value and the voltage deviation exceeds the threshold value, the control needs to be triggered. The third step is fuzzy reasoning, and the rule weight is calculated through neural network forward propagation, and the reasoning result is integrated. The fourth step is defuzzification, which converts the fuzzy result into a clear conclusion of whether the threshold value is exceeded or not, and finally realizes the joint judgment of multiple variables. After presetting the power fluctuation threshold value, the voltage stability threshold value and the frequency stability threshold value, when the real-time power difference exceeds the threshold value and the voltage deviation and the frequency deviation are normal, the ANFIS outputs the dynamic cooperative control instruction that needs to be triggered by matching the corresponding rules. When the power difference, the voltage deviation and the frequency deviation are all normal, the ANFIS outputs the current state. Thus, the one-sidedness problem of the traditional single parameter threshold judgment is solved, such as false triggering caused by only looking at the power difference exceeding the threshold value while ignoring the normal voltage and frequency. At the same time, the rule base is continuously optimized through neural network self-learning to adapt to the change of microgrid working conditions, such as load fluctuation caused by seasonal replacement, to improve the accuracy of threshold judgment and reduce the false triggering rate.

[0035] As an optional embodiment, in step 104, a preset adaptive neuro-fuzzy inference system ANFIS is called to take the real-time power difference value, the busbar voltage deviation value and the frequency deviation value as input variables, and to determine whether the real-time power difference value exceeds the preset power fluctuation threshold value and whether the busbar voltage deviation value or the frequency deviation value exceeds the preset stability threshold value through the fuzzy rule base and the neural network training model of the ANFIS, including: adopting a fuzzy C-means clustering FCM algorithm to perform clustering analysis on the real-time power difference value, the busbar voltage deviation value and the frequency deviation value data in the historical operation of the microgrid to generate initial fuzzy subsets and membership functions of the ANFIS fuzzy rule base; taking the threshold-exceeding samples and normal samples in the historical data as a training set to train the ANFIS, and optimizing the membership function parameters and the neural network connection weights through a back propagation algorithm to make the identification accuracy of the model to the threshold-exceeding samples meet preset requirements; calling the trained ANFIS to input the real-time calculated microgrid real-time power difference value, busbar voltage deviation value and frequency deviation value into the model; performing fuzzy inference through the fuzzy rule base to output fuzzy judgment results corresponding to each input variable; and converting the fuzzy judgment results into clear numerical judgment conclusions through the de-fuzzification processing of the ANFIS to determine whether the real-time power difference value exceeds the preset power fluctuation threshold value and whether the busbar voltage deviation value or the frequency deviation value exceeds the preset stability threshold value.

[0036] In the ANFIS processing process involved in step 104, the FCM algorithm first needs to preprocess the real-time power difference value, busbar voltage deviation value and frequency deviation value data in the historical operation of the microgrid collected to remove abnormal values and initialize K clustering centers (the number of clusters is determined according to the category characteristics of the parameter states in the historical data, such as normal state, power difference value exceeding state, voltage deviation exceeding state and frequency deviation exceeding state), and then iteratively calculate the membership of each historical sample to each clustering center. The distance between the sample and the clustering center is calculated through the defined membership function, and the membership is higher when the distance is closer. Meanwhile, the objective function of FCM (with the minimization of the weighted square error of the membership as the target) is constructed, and the objective function is optimized by constantly updating the membership matrix and the clustering center until the change amount of the clustering center in the adjacent two iterations is less than a preset threshold value. At this time, each cluster corresponds to a fuzzy subset, and the membership function parameters (such as the mean and standard deviation of the Gaussian membership function, the mean corresponding to the clustering center and the standard deviation corresponding to the half-width of the clustering range) of each input variable (real-time power difference value, busbar voltage deviation value and frequency deviation value) can be determined based on the clustering center and the clustering range, so as to generate the initial fuzzy subsets and membership functions of the ANFIS fuzzy rule base.

[0037] It can be further understood that the fuzzy C-means clustering (FCM) algorithm is an unsupervised clustering method based on fuzzy set theory, and the core idea is to avoid the limitations of traditional hard clustering (each sample belongs to only one category), allowing samples to belong to multiple clusters with different fuzzy membership degrees (value range [0, 1]), which is more suitable for the case of fuzzy sample category boundary in real scenarios.

[0038] Next, the ANFIS training phase, the core calculation is around the back propagation algorithm, first of all, the threshold of the historical data samples (corresponding to the expected output is to trigger control) and normal samples (corresponding to the expected output is to maintain state) are divided into training set, the training process needs to combine the hierarchical structure of ANFIS to carry out forward propagation and back propagation calculation. When forward propagation, the fuzzification layer converts the input historical sample parameters into fuzzy membership values through the generated membership function, the rule layer calculates the trigger strength of each fuzzy rule (usually the product or maximum operation of the membership values of each input variable under the same rule), the normalization layer normalizes the trigger strength of each rule (the normalized trigger strength of a rule is the ratio of the trigger strength of the rule to the sum of the trigger strengths of all rules), and the output layer calculates the predicted output of ANFIS according to the normalized trigger strength and rule consequent parameters (initially set randomly). Then, by comparing the predicted output with the expected output of the sample, the error (such as mean square error) is calculated, and the back propagation phase is entered, the error is transferred along the network level in reverse, the membership function parameters (such as the mean and standard deviation of the Gaussian function) and the neural network connection weights (mainly the linear parameters of the rule consequent) are adjusted by gradient descent method, these parameters are updated in each iteration to reduce the error, and the training is repeated until the model's threshold exceeding identification accuracy of the training set meets the preset requirements, at this time, the fuzzy rule base of ANFIS and the neural network parameters have been optimized.

[0039] When the trained ANFIS is invoked for real-time judgment, the core calculation begins with the fuzzification of real-time parameters. The real-time calculated power difference, bus voltage deviation, and frequency deviation of the microgrid are input into the corresponding fuzzification layers. The membership function, which is optimized, is used to calculate the membership value of each parameter to each fuzzy subset (e.g., the degree to which the real-time power difference belongs to the fuzzy subsets such as "normal," "slightly exceeding the limit," and "severely exceeding the limit"). Subsequently, fuzzy inference calculations are performed at the rule layer. All rules in the fuzzy rule base are traversed, and the trigger strength of each rule is calculated based on the membership value of each input parameter. For example, if a rule is "if the real-time power difference exceeds the limit and the bus voltage deviation and frequency deviation are normal, then the power difference is determined to exceed the threshold," then the membership value of the "exceeding the limit" subset corresponding to the real-time power difference is multiplied by the membership values ​​of the "normal" subset corresponding to the bus voltage deviation and the "normal" subset corresponding to the frequency deviation to obtain the trigger strength of the rule. Next, the trigger strength of all rules is normalized in the normalization layer to determine the contribution weight of each rule in the overall reasoning result. Then, in the output layer, the normalized trigger strength of each rule is multiplied by the output value of the rule consequent (such as the value 1 corresponding to "exceeding the threshold" and the value 0 corresponding to "normal") and summed to obtain the fuzzy judgment result corresponding to each input variable (such as the fuzzy output corresponding to the power difference, the fuzzy output corresponding to the voltage deviation, and the fuzzy output corresponding to the frequency deviation).

[0040] Finally, defuzzification calculation is performed. Since the output of fuzzy inference is a continuous fuzzy value, it needs to be transformed into a clear numerical judgment conclusion through a defuzzification algorithm. The centroid method calculation process is as follows: based on the numerical range of the fuzzy judgment result, calculate the sum of the products of all possible values ​​in the range and the corresponding membership values, and then divide by the sum of the membership values ​​to obtain a clear value. If the value is greater than the preset judgment threshold (such as 0.5), it is determined that the corresponding input variable exceeds the threshold; otherwise, it is determined that it does not exceed the threshold. Thus, it is determined whether the real-time power difference exceeds the preset power fluctuation threshold and whether the bus voltage deviation or frequency deviation exceeds the preset stability threshold.

[0041] Step 105 is to control instruction execution and state maintenance, which is based on the judgment result of ANFIS to execute differentiated control strategy. When all parameters are not over the threshold value, the current state of the microturbine and the energy storage is maintained to avoid unnecessary adjustment leading to equipment loss. When any parameter is over the threshold value, the dynamic cooperative control instruction is triggered to start the subsequent energy storage fast response and microturbine cooperative adjustment to ensure that the microgrid quickly returns to a stable state, and the core logic is to save energy when stable and control when fluctuating. When ANFIS judges that all parameters are normal, the community microgrid maintains the output of the microturbine and the regular charging of the energy storage system. When ANFIS judges that the power difference is over the threshold value and negative, the dynamic cooperative control instruction is immediately triggered to control the energy storage system to increase the discharge power to make up for the gap, and at the same time the microturbine is instructed to gradually increase the output. When the voltage deviation of the industrial park microgrid is over the threshold value, the instruction is triggered to control the energy storage system to adjust the charging and discharging power to stabilize the voltage, and at the same time the fuel supply of the microturbine is adjusted to optimize the output. In this way, the stability and economy of the microgrid are balanced, the microturbine is prevented from frequent start-stop and the energy storage is prevented from frequent charging and discharging, the fuel consumption and mechanical wear of the microturbine are reduced, the cycle loss of the energy storage battery is reduced, and the service life of the equipment is prolonged. When fluctuating, the control is triggered quickly to prevent power imbalance from expanding into voltage collapse and frequency instability, to avoid load tripping caused by low frequency, and to improve the reliability of microgrid power supply.

[0042] As an optional embodiment, in step 105, after the dynamic cooperative control instruction is triggered, it further includes: preferentially controlling the energy storage system to perform millisecond / second response adjustment, determining the power adjustment direction and adjustment amount based on the real-time power difference; when the real-time power difference is positive, it is determined that the microgrid is in an energy surplus state, and the energy storage system is controlled to increase the charging power to absorb the excess power; when the real-time power difference is negative, it is determined that the microgrid is in an energy deficiency state, and the energy storage system is controlled to increase the discharging power to make up for the power gap; a model predictive control (MPC) algorithm is used to predict the future short-term power fluctuation according to the change trend of the real-time power difference, dynamically correct the charging and discharging power instruction of the energy storage system, and suppress the short-term fluctuation caused by sudden drop of photovoltaic / wind power output or sudden increase of load.

[0043] In the above steps, the core calculation of the dynamic cooperative control instruction triggered to preferentially control the energy storage system to respond to the adjustment begins with the analysis of the real-time power difference and the determination of the adjustment parameter. First, according to the positive and negative of the real-time power difference, the energy supply and demand state of the microgrid is determined. When the real-time power difference is positive, the difference is directly taken as the basic quantitative index of the energy surplus, and the adjustment amount of the charging power of the energy storage system to be improved is calculated. At this time, the current charging and discharging state parameters of the energy storage system (such as the current charging power, the remaining capacity, and the maximum charging power limit) need to be combined, and the smaller value between the real-time power difference and the maximum charging power that can be improved by the energy storage system (the maximum charging power minus the current charging power) is taken as the final charging power adjustment amount to avoid exceeding the physical operation limit of the energy storage system. When the real-time power difference is negative, the absolute value of the difference is taken as the basic quantitative index of the energy deficiency, and the adjustment amount of the discharging power of the energy storage system to be improved is calculated in the same way by combining the current discharging power, the remaining capacity, and the maximum discharging power limit of the energy storage system. That is, the smaller value between the absolute value of the real-time power difference and the maximum discharging power that can be improved by the energy storage system (the maximum discharging power minus the current discharging power) is taken to determine the final discharging power adjustment direction and adjustment amount, and then the initial charging and discharging power instruction of the energy storage system is generated.

[0044] The core calculation of dynamic correction using model predictive control (MPC) algorithm needs to construct an MPC prediction model and determine input variables first. Real-time power difference, real-time output of photovoltaic power source, real-time output of wind power source and real-time demand data of controllable load are obtained through the microgrid real-time monitoring module after preprocessing, and the change rate of each variable (the difference between the current time data and the previous sampling time data divided by the sampling time interval) is calculated. The real-time power difference and its change rate, the change rate of photovoltaic output, the change rate of wind power output and the change rate of controllable load demand are used as input variables of the MPC prediction model. Then, the MPC prediction model predicts the microgrid power fluctuation in the short term (such as 10 seconds to 1 minute in the future, corresponding to the response period of the energy storage system) based on the power fluctuation law learned in the historical training process (obtained by training the power difference time series data, photovoltaic and wind power output fluctuation data and load demand fluctuation data in the historical operation of the microgrid), and outputs the power fluctuation prediction curve in the short term in the future through the rolling optimization calculation inside the model. The curve needs to reflect the predicted value of the real-time power difference at different times; according to the prediction curve, the trend of future power difference is analyzed, if the prediction curve shows that the future real-time power difference will change in the increasing direction (energy surplus or energy shortage will intensify), the positive correction amount (increase the charging or discharging power) that needs to be superimposed is calculated, and the size of the correction amount is determined according to the deviation degree of the predicted difference and the current difference, the larger the deviation, the larger the correction amount; if the prediction curve shows that the future real-time power difference will change in the gentle direction (tend to zero), the reverse correction amount (reduce the charging or discharging power) that needs to be superimposed is calculated. Finally, the charging and discharging power correction amount calculated is superimposed with the initial charging and discharging power instruction of the energy storage system, if it is charging adjustment, the initial charging power instruction is added with the charging correction amount, if it is discharging adjustment, the initial discharging power instruction is added with the discharging correction amount, to generate the final charging and discharging power control instruction of the energy storage system, which controls the operation of the energy storage system through the instruction, and realizes the suppression of short-term fluctuation caused by sudden drop of photovoltaic and wind power output or sudden increase of load.

[0045] As an optional embodiment, in the above step, the model predictive control (MPC) algorithm is used to predict future short-term power fluctuation according to the trend of real-time power difference, and dynamically correct the charge and discharge power instruction of the energy storage system to suppress short-term fluctuation caused by sudden drop of photovoltaic / wind power output or sudden increase of load, comprising: constructing a prediction model of the MPC algorithm, taking the real-time power difference of the microgrid, the real-time output change rate of the photovoltaic power source, the real-time output change rate of the wind power source, and the real-time demand change rate of the controllable load as the input variables of the model, the training samples of the prediction model are the power difference time series data, the photovoltaic / wind power fluctuation data and the load demand fluctuation data in the historical operation of the microgrid, the model parameters are optimized through offline training to make the prediction accuracy of the model meet the prediction demand of short-term power fluctuation; based on the preprocessed data output by the microgrid real-time monitoring module, the real-time power difference, the real-time output change rate of the photovoltaic power source, the real-time output change rate of the wind power source, and the real-time demand change rate of the controllable load are obtained in real time and input into the trained MPC prediction model, and the future short-term microgrid power fluctuation prediction curve is output; according to the power fluctuation prediction curve, the theoretical charge and discharge power correction amount required by the energy storage system is calculated; if the prediction curve shows that the power difference will increase, the charge power correction amount or the discharge power correction amount of the energy storage system is increased; if the prediction curve shows that the power difference will tend to be flat, the charge power correction amount or the discharge power correction amount of the energy storage system is reduced; the charge and discharge power correction amount is superimposed on the initial charge and discharge power instruction of the energy storage system to generate the final charge and discharge power control instruction of the energy storage system, and the energy storage system is controlled to operate according to the instruction to suppress short-term fluctuation caused by sudden drop of photovoltaic / wind power output or sudden increase of load; during the operation of the MPC algorithm, the latest real-time power difference collected by the microgrid real-time monitoring module is continuously compared with the prediction value of the MPC prediction model, the prediction deviation is calculated, and if the deviation exceeds the preset deviation threshold, the parameters of the MPC prediction model are corrected online to ensure the accuracy of subsequent power fluctuation prediction and maintain the stability of the adjustment effect of the energy storage system.

[0046] In the above step, in the construction of the input variables of the MPC algorithm prediction model, the microgrid real-time power difference, the photovoltaic power source real-time output change rate, the wind power source real-time output change rate, and the controllable load real-time demand change rate are determined as the input variables of the model, wherein each output change rate and demand change rate is calculated by dividing the difference between the current time data and the previous sampling time data by the sampling time interval, and the training samples are extracted from the power difference time series data, the photovoltaic / wind power fluctuation data, and the load demand fluctuation data in the historical operation data of the microgrid. After standardizing these samples to eliminate the difference in variable magnitude, the gradient descent optimization algorithm is used for offline training, the parameters such as weights and biases in the model are adjusted repeatedly, the error between the model prediction output and the historical actual power fluctuation data is gradually reduced, and the prediction accuracy of the model meets the prediction demand of short-term power fluctuation.

[0047] Based on the preprocessed data output by the microgrid real-time monitoring module, the real-time input variables are obtained, and the real-time power difference, the real-time photovoltaic output, the real-time wind power output and the real-time controllable load demand data are dynamically calculated. The real-time power difference is directly extracted, and the photovoltaic output rate, the wind power output rate and the controllable load demand rate are calculated respectively. After inputting these real-time input variables into the trained MPC prediction model, the model predicts the microgrid power difference at different times in the future short term through the internal preset time sequence prediction logic and the power fluctuation law learned in the training process. The prediction values are integrated in time sequence to generate a microgrid power fluctuation prediction curve in the future short term.

[0048] When the power fluctuation prediction curve is used to calculate the theoretical charge and discharge power correction amount of the energy storage system, the trend of the prediction curve is analyzed, the change direction and amplitude of the power difference are determined by calculating the slope of the prediction curve at each time, and if the slope is positive and the absolute value increases, it indicates that the power difference will increase. At this time, the correction direction is determined according to the positive and negative of the power difference (positive for energy surplus and negative for energy deficiency), and the charging power correction amount or the discharging power correction amount of the energy storage system is increased. The size of the correction amount is determined according to the absolute value of the slope, and the larger the absolute value of the slope, the larger the correction amount. If the slope tends to zero, it indicates that the power difference will tend to be flat, and the charging power correction amount or the discharging power correction amount of the energy storage system is reduced. Similarly, the size of the correction amount is adjusted according to the change amplitude of the slope.

[0049] When the charging and discharging power correction amount is added to the initial charging and discharging power instruction of the energy storage system, the superposition logic is determined according to the positive and negative of the power difference. If the initial instruction is a charging instruction and the power difference is positive, the charging power correction amount is added to the initial charging power instruction to obtain the final charging power control instruction. If the initial instruction is a discharging instruction and the power difference is negative, the discharging power correction amount is added to the initial discharging power instruction to obtain the final discharging power control instruction. The energy storage system is controlled by the final instruction to suppress the short-term fluctuation caused by the sudden drop of photovoltaic and wind power output or the sudden increase of load.

[0050] During the operation of the MPC algorithm, the core calculation of online correction is to continuously subtract the latest real-time power difference collected by the microgrid real-time monitoring module from the prediction value at the corresponding time of the MPC prediction model to obtain the prediction deviation. The deviation is compared with the preset deviation threshold. If the deviation exceeds the threshold, the latest real-time power difference, the real-time change rate and the prediction deviation are used as the basis to fine-tune the weight deviation and other parameters of the MPC prediction model through algorithms such as back propagation, so that the model can adapt to the changes of the current microgrid operating conditions, ensure the accuracy of subsequent power fluctuation prediction, and maintain the stability of the adjustment effect of the energy storage system.

[0051] As an optional embodiment, after the dynamic cooperative control instruction is triggered in step 105, the method further comprises: based on the real-time power difference value and the current operation mode of the microgrid, generating a gas turbine output adjustment instruction in combination with a long short-term memory network (LSTM) prediction model when the energy storage system starts to respond to the adjustment; the LSTM prediction model takes historical power difference data, photovoltaic / wind power output prediction values and load prediction values as training samples, and outputs a future medium and long term microgrid power demand prediction result in real time; if the microgrid is in a grid-connected mode, the fuel supply amount of the gas turbine is adjusted to gradually adjust the gas turbine output in combination with the grid interaction power limit value and the microgrid power demand prediction result, so that the gas turbine output and the energy supplement / absorption effect of the energy storage system form a time sequence cooperation, and the real-time power difference value of the microgrid is stably controlled within a preset range; if the microgrid is in an off-grid mode, the starting time sequence or the output increase range of the gas turbine is adjusted based on the remaining capacity of the energy storage system and the discharge duration of the energy storage system predicted by the LSTM, so that the gas turbine output completes the energy supplement connection before the response of the energy storage system decays, and transient imbalance of the microgrid is avoided; the real-time monitoring module is continuously used to track the changes of the microgrid bus voltage, frequency and power difference value, and when it is monitored that the microgrid bus voltage and frequency continuously return to the preset stable threshold value, and the real-time power difference value continuously maintains within the preset power fluctuation threshold value, a PID closed-loop adjustment algorithm is called to control the energy storage system to reduce the charging and discharging power to a normal charging and discharging state, so that the gas turbine output is stably controlled at the current adjusted operating point, and a dynamic cooperative control process is completed.

[0052] In the above steps, the core calculation of generating a gas turbine output adjustment instruction in combination with a long short-term memory network (LSTM) prediction model when the energy storage system starts to respond to the adjustment first involves training and real-time prediction of the LSTM prediction model. In the training stage, historical power difference data, photovoltaic / wind power output prediction values and load prediction values are processed in time sequence, and these data are divided into input sequences and corresponding output sequences (the output sequences are actual values of microgrid power demand in a future period) according to a fixed time window. The weights of the forget gate, input gate and output gate of the LSTM model and the cell state parameters are iteratively optimized through a back propagation algorithm, so that the model can learn the time sequence correlation rules in the historical data. In real-time prediction, the latest historical power difference data, current photovoltaic / wind power output prediction values and current load prediction values output by the real-time monitoring module of the microgrid are obtained, a real-time input sequence is constructed according to the time window format in the training, and the trained LSTM model is input. The model filters historical key information through the forget gate, updates the cell state through the input gate, and generates the current prediction value through the output gate, and finally outputs a future medium and long term microgrid power demand prediction result.

[0053] If the microgrid is in grid-connected mode, the core calculation needs to first determine the upper and lower boundaries of the grid interaction power limit, compare it with the microgrid power demand prediction result output by LSTM, and calculate the target output range of the gas turbine under the premise of meeting the power demand and not exceeding the grid interaction power limit. First, subtract the current photovoltaic and wind power output prediction value from the power demand prediction result to get the preliminary gas turbine output demand. Then, adjust this output demand in combination with the grid interaction power limit. If the preliminary output demand leads to the interaction power between the microgrid and the grid exceeding the limit, then correct the gas turbine output demand in the reverse direction according to the limit. After determining the target output of the gas turbine, calculate the adjustment amount of the fuel supply based on the corresponding relationship between the gas turbine output and the fuel supply (a function relationship fitted through historical operation data of the gas turbine). At the same time, combine the current charging or discharging power and the trend of the energy storage system to calculate the adjustment rate of the gas turbine output (to ensure that the change of the gas turbine output is synchronized in time with the response of the energy storage system). By gradually adjusting the fuel supply, the gas turbine output approaches the target value, and the real-time power difference is stabilized within the preset range.

[0054] If the microgrid is in off-grid mode, analyze the remaining capacity and current discharge power of the energy storage system, divide the remaining capacity by the current discharge power to get the current discharge duration of the energy storage system, compare it with the LSTM predicted discharge duration of the energy storage system, and determine the time node of the response decay of the energy storage system. At the same time, take the stability of the microgrid bus voltage and frequency as the target, calculate the amount of gas turbine output needed to make the voltage and frequency return to stability through the correlation model between voltage and frequency and gas turbine output (such as the adjustment relationship between voltage deviation and active power, the adjustment relationship between frequency deviation and active power), and determine the starting time sequence of the gas turbine (if the discharge duration of the energy storage system is short, the delay time of the gas turbine startup is shortened) or the output increase amplitude (if the power gap needs to be quickly filled, the output increase amplitude of the gas turbine is increased) in combination with the discharge duration of the energy storage system, to ensure that the gas turbine output can reach the required supplement before the response of the energy storage system decays.

[0055] The core calculation in the process of continuously tracking the microgrid parameters is to compare the monitored bus voltage, frequency and power difference with the preset threshold in real time. When these parameters are continuously within the threshold, the PID closed-loop adjustment algorithm is called. First, determine the target power of the energy storage system in the normal charging and discharging state, calculate the deviation value of the current energy storage charging and discharging power from the target power, and use the deviation value as the input of the PID algorithm to calculate the adjustment amount of the proportional link (deviation value multiplied by proportional coefficient), integral link (deviation value integral multiplied by integral coefficient) and differential link (deviation value change rate multiplied by differential coefficient). Add the three adjustment amounts to get the total adjustment amount of the energy storage charging and discharging power, generate a control instruction to gradually reduce the charging and discharging power of the energy storage system to the normal state according to the adjustment amount, and keep the current output of the gas turbine unchanged until the energy storage system stabilizes in the normal charging and discharging state, completing a dynamic cooperative control.

[0056] As an optional embodiment, in the above step, the generating of the gas turbine output adjustment instruction based on the real-time power difference value and the current operation mode of the microgrid, in combination with the long short-term memory network (LSTM) prediction model, comprises: constructing an LSTM prediction model architecture integrating an attention mechanism, the architecture including an input layer, an attention layer, a bidirectional LSTM hidden layer, a regularization layer, and an output layer; wherein the input layer receives historical power difference data of the microgrid, predicted values of photovoltaic power output, predicted values of wind power output, and predicted values of controllable load demand; the attention layer strengthens the feature expression of key variables by calculating the contribution weight of each input variable to the power demand of the microgrid; the bidirectional LSTM hidden layer is used to capture the bidirectional dependency relationship of the input time series data, and the regularization layer suppresses model overfitting through an adaptive weight decay mechanism; a federated learning framework is used to perform distributed training on the LSTM prediction model, historical operation data of multiple microgrids of the same type are used as distributed training samples, the model gradient of each microgrid is aggregated through a federated averaging algorithm, the model parameter collaborative optimization across microgrids is realized, and the adaptability of the model to different microgrid scenarios is improved; during the training process, the mean absolute percentage error (MAPE) is used as the loss function, and the training step is dynamically adjusted through an adaptive learning rate optimization algorithm until the prediction accuracy of the model meets the preset requirements; the input data of the model, including historical power difference time series data, predicted values of photovoltaic / wind power output, and predicted values of controllable load demand, are obtained through a microgrid real-time monitoring module, and after standardization preprocessing, the input data are input into the trained LSTM prediction model; the LSTM prediction model dynamically allocates variable weights through the attention layer, performs time series feature extraction through the bidirectional LSTM hidden layer, and finally outputs the predicted results of the medium and long term power demand of the microgrid in the future; the predicted results of the power demand are dynamically corrected in combination with the current operation mode of the microgrid and the real-time power difference value; in the grid-connected mode, a large grid interactive power constraint factor is introduced, the predicted results are adjusted according to the grid dispatching instruction, and it is ensured that the gas turbine output adjustment conforms to the grid interactive power limit value; in the off-grid mode, an energy storage system state constraint factor is introduced, the predicted results are corrected based on the remaining capacity and charge / discharge characteristics of the energy storage system, and the time sequence mismatch between the gas turbine output and the energy storage compensation is avoided; based on the corrected predicted results of the power demand, in combination with the operating characteristics of the gas turbine, the gas turbine output adjustment instruction is generated; the instruction includes a gas turbine target output reference value, an output adjustment gradient, and a start / stop time sequence logic, so as to ensure that the gas turbine output adjustment can respond to the change of the power demand of the microgrid and maintain the stability of the gas turbine itself, and at the same time, the compensation / energy absorption effect of the energy storage system is formed in cooperation.

[0057] In the above steps, when constructing the LSTM prediction model architecture with attention mechanism, the input layer quantifies the functions of each layer to convert the historical power difference data of the microgrid, the predicted output of the photovoltaic power supply, the predicted output of the wind power supply, and the predicted demand of the controllable load into a unified dimensional vector form, which is mapped to the feature space that the model can process through matrix operations. The attention layer determines the contribution weight by calculating the correlation between each input variable and the power demand of the microgrid. Specifically, the input variables are first linearly transformed to generate a query key matrix, and then the original attention weight is obtained through the dot product operation of query and key. After normalization by the softmax function, the weighted feature vector is output by multiplying the value matrix. The feature expression of the key variable is strengthened. The bidirectional LSTM hidden layer is composed of a forward LSTM and a backward LSTM. The forward LSTM processes the time series data in chronological order to capture future dependencies, and the backward LSTM processes the data in reverse chronological order to capture past dependencies. The outputs of the two are integrated into a feature vector containing bidirectional time series information by splicing. The regularization layer suppresses model overfitting by calculating the L2 norm of each parameter and incorporating it into the loss function to achieve adaptive weight decay. The output layer maps the hidden layer output to the future medium and long-term microgrid power demand prediction value through a fully connected network.

[0058] Distributed training using the federated learning framework involves first dividing the historical operation data of multiple microgrids of the same type into local training samples. Each microgrid independently trains the model based on the local samples, calculates the parameter gradient, and uploads only the gradient to the central server. The central server weights and averages the gradients uploaded by each microgrid based on the sample size to obtain the global gradient, which is then distributed to each microgrid to update the local model parameters. This iteration continues until the model converges. During training, the mean absolute percentage error (MAPE) is used as the loss function to measure the prediction accuracy by calculating the average proportion of the absolute error of the predicted value to the actual value. An adaptive learning rate optimization algorithm is also used to dynamically adjust the training step size based on the rate of change of the loss function. When the loss decreases rapidly, the step size is increased to accelerate convergence. When the loss fluctuates, the step size is reduced to ensure stability until the model prediction accuracy meets the preset requirements.

[0059] When obtaining model input data and processing, the historical power difference time series data, photovoltaic and wind power output prediction values, and controllable load demand prediction values are extracted from the microgrid real-time monitoring module. The data is standardized by calculating the mean and standard deviation of each variable to convert it to the same order of magnitude in the feature space, eliminating the dimension effect. The processed data is input into the trained LSTM prediction model. The model first recalculates the contribution weight of the real-time input variables through the attention layer to strengthen the influence of the current key variables (such as the fluctuating photovoltaic output prediction value). Then, the standardized time series data is feature extracted by the bidirectional LSTM hidden layer to fuse past and future dependencies. Finally, the future medium and long-term microgrid power demand prediction result is output by the output layer.

[0060] When the current operation mode of the microgrid and the real-time power difference are combined to dynamically correct the prediction results, different constraint factors need to be introduced according to the mode difference. In the grid-connected mode, the interactive power limit value in the grid dispatching instruction is obtained first, the deviation between the limit value and the current microgrid predicted power is calculated, the deviation is quantified as the grid interactive power constraint factor, and the power demand prediction result is adjusted through multiplication operation. The corrected prediction value corresponds to the gas turbine output adjustment that does not exceed the grid interactive power limit value. In the off-grid mode, based on the remaining capacity of the energy storage system and the charge-discharge characteristic curve, the power adjustment duration supported by the energy storage is calculated. The duration is compared with the predicted power demand duration to quantify the energy storage system state constraint factor. The prediction result is also corrected through multiplication operation to avoid the mismatch between the gas turbine output and the energy storage compensation in time.

[0061] Based on the corrected power demand prediction result, the gas turbine output adjustment instruction is generated, which combines the operation characteristics of the gas turbine. First, the corrected power demand prediction result is subtracted from the current photovoltaic and wind power output prediction value to obtain the target output reference value that the gas turbine needs to bear. Then, the output adjustment gradient is calculated according to the output adjustment rate limit of the gas turbine (determined by the gas turbine model and operation parameters) to ensure that the output change per unit time does not exceed the safety threshold. At the same time, the start-up or shutdown timing logic of the gas turbine is determined according to the time sequence of the power demand prediction. If the predicted future power demand is higher than the current output, the start-up timing or the output gradient is set. If the predicted demand decreases, the shutdown timing or the output gradient is set. Finally, the target output reference value, the output adjustment gradient, and the start-up and shutdown timing logic are integrated to form a complete gas turbine output adjustment instruction, which ensures that the gas turbine adjustment responds to the microgrid power demand change and maintains the stability of its own operation, and the compensation and absorption effect of the energy storage system forms a synergy.

[0062] As an optional embodiment, in the above steps, the microgrid bus voltage, frequency and power difference value changes are continuously tracked by the real-time monitoring module. When it is monitored that the microgrid bus voltage and frequency continuously return to the preset stable threshold, and the real-time power difference continuously maintains within the preset power fluctuation threshold, a PID closed-loop regulation algorithm is called to control the energy storage system to reduce the charging and discharging power to the normal charging and discharging state, including: a fuzzy PID regulation model integrating reinforcement learning is constructed, the model includes a fuzzy reasoning unit and a reinforcement learning optimization unit; wherein the fuzzy reasoning unit takes the microgrid bus voltage deviation and deviation change rate, the frequency deviation and deviation change rate, and the real-time power difference deviation and deviation change rate as inputs, and maps the input variables to the adjustment interval of the PID parameters through a preset fuzzy rule base; the reinforcement learning optimization unit takes the smooth transition of the energy storage system charging and discharging power to the normal state and the non-overshoot fluctuation of the microgrid parameters as the reward function, dynamically optimizes the PID parameter adjustment interval output by the fuzzy reasoning unit through real-time interaction with the microgrid operating environment, and improves the parameter adaptability; the real-time data of the microgrid bus voltage, frequency and real-time power difference are obtained through the microgrid real-time monitoring module, the deviations and deviation change rates of the parameters from the preset threshold are calculated, and are input into the fuzzy PID regulation model integrating reinforcement learning; after the model is optimized by fuzzy reasoning and reinforcement learning, the optimal PID parameters under the current working condition are output. Based on the output PID parameters, a charging and discharging power regulation closed loop of the energy storage system is constructed; the difference between the current charging and discharging power of the energy storage system and the normal charging and discharging power is taken as the PID regulation deviation, the charging and discharging power regulation amount of the energy storage system is calculated through the PID algorithm, and the energy storage system power control instruction is generated. At the same time, a microgrid parameter feedback mechanism is introduced. If the microgrid bus voltage, frequency or real-time power difference deviates from the preset threshold during the adjustment process, the PID regulation deviation is corrected in real time to avoid fluctuations in the microgrid parameters caused by the adjustment process. The energy storage system gradually reduces the charging and discharging power according to the power control instruction, and simultaneously tracks the charging and discharging power change of the energy storage system and the microgrid parameter stability through the microgrid real-time monitoring module. If the charging and discharging power of the energy storage system approaches the normal charging and discharging power range, a PID parameter fine-tuning mode is triggered, the adjustment rate is reduced by reducing the proportional coefficient and the integral coefficient, and the smooth transition of the energy storage system power to the normal state is ensured. When the microgrid real-time monitoring module monitors that the charging and discharging power of the energy storage system continuously and stably maintains within the normal charging and discharging power range, and the microgrid bus voltage, frequency and real-time power difference all continuously maintain within the preset threshold, it is determined that the energy storage system regulation is completed, the PID closed-loop regulation algorithm is stopped, and the current energy storage system operating state is maintained to provide a basic condition for the next dynamic cooperative control.

[0063] In the above steps, the fuzzy inference unit and the reinforcement learning optimization unit are designed in cooperation. The fuzzy inference unit first quantitatively processes the microgrid bus voltage deviation and the deviation change rate, the frequency deviation and the deviation change rate, and the real-time power difference deviation and the deviation change rate. The voltage deviation is the difference between the real-time voltage and the preset threshold, and the deviation change rate is the difference between the current deviation and the deviation at the previous moment divided by the sampling time. The deviation and the change rate of the frequency and the power difference are calculated in the same way. These input variables are converted into fuzzy language variables such as positive large, positive medium, negative small, etc. through fuzzy processing. The fuzzy variables are mapped to the preliminary adjustment interval of the PID parameters (proportion coefficient, integral coefficient, and differential coefficient) through a preset fuzzy rule base (the rules are based on the microgrid regulation experience, such as increasing the proportion coefficient when the voltage deviation is large and the change rate is large). The reinforcement learning optimization unit constructs a reward function based on the degree of smooth transition of the energy storage system charging and discharging power to the normal state (the smaller the power change rate, the greater the reward) and the condition that the microgrid parameters have no overshoot and fluctuation (the smaller the amplitude of the parameter deviation from the threshold, the greater the reward). Through real-time interaction with the microgrid operating environment (the microgrid state and the reward are obtained after each adjustment), the value function is updated using the time difference algorithm to dynamically optimize the PID parameter adjustment interval output by the fuzzy inference unit, so that the parameters are more suitable for the current working condition.

[0064] After obtaining the data through the microgrid real-time monitoring module, the core calculation focuses on the real-time solution of the deviation and the change rate. The bus voltage frequency and the real-time power difference are extracted from the monitoring data, and the deviation from the corresponding preset threshold is calculated. Then, the deviation change rate of each parameter is obtained by dividing the deviation difference between adjacent moments by the sampling interval. These calculation results are input into the fuzzy PID regulation model based on reinforcement learning. The model first performs fuzzy operation on the input deviation and change rate through the fuzzy inference unit, and outputs the initial adjustment interval of the PID parameters according to the fuzzy rule base. The reinforcement learning optimization unit modifies the initial interval based on the current microgrid operating state (such as the current power of the energy storage system and the stability of the microgrid parameters). The optimal PID parameter values (specific values of the proportion coefficient, the integral coefficient, and the differential coefficient) are determined by maximizing the cumulative reward.

[0065] Based on the output PID parameters, a closed-loop power regulation of the energy storage system is constructed. The difference between the current charging and discharging power of the energy storage system and the normal charging and discharging power is taken as the PID regulation deviation. The proportioning part directly multiplies the deviation by the proportioning coefficient to obtain the proportioning adjustment amount. The integral part integrates the deviation over time and then multiplies it by the integral coefficient to obtain the integral adjustment amount. The differential part calculates the change rate of the deviation (the difference between the current deviation and the deviation at the previous moment divided by the sampling time) and then multiplies it by the differential coefficient to obtain the differential adjustment amount. The sum of the three is the total adjustment amount of the charging and discharging power of the energy storage system, from which the power control command is generated.

[0066] The core calculation of introducing the microgrid parameter feedback mechanism is to continuously monitor the bus voltage frequency and real-time power difference in the adjustment process. If it is detected that a parameter deviates from the preset threshold (determined by calculating the deviation rate of the parameter, which is positive and the absolute value increases), the PID adjustment deviation is dynamically corrected according to the deviation degree. For example, when the voltage has a deviation trend, increase the weight of the voltage deviation in the total adjustment deviation. Adjust the proportional integral derivative adjustment value indirectly to avoid the adjustment process causing the microgrid parameter fluctuation.

[0067] When the monitoring of the energy storage system charging and discharging power is close to the normal range (the difference between the current power and the normal power is less than the preset proportion), the PID parameter fine-tuning mode is triggered, the adjustment rate is reduced by reducing the proportional coefficient and the integral coefficient. The specific calculation is to reduce the original proportional coefficient and integral coefficient value by a preset proportion, so that the adjustment amount is more gentle as the deviation decreases, and the power is smoothly transitioned to the normal state.

[0068] In the embodiments of the present application, through data preprocessing and ANFIS optimization, the accuracy and timeliness of microgrid operation state judgment are greatly improved, and the problems of data noise interference and one-sided judgment in traditional control are solved. Through the differential instruction triggering strategy, while ensuring the stability of microgrid voltage, frequency and power supply and demand balance, the invalid adjustment loss of equipment is reduced, and the service life of gas turbine and energy storage system is prolonged. Ultimately, the microgrid operation stability, power supply reliability and operation efficiency are synergistically improved, which provides key technical support for the efficient cooperation of gas turbine and microgrid in different scenarios such as community and industrial park, especially adapts to the complex operation demand of microgrid after the access of renewable energy such as photovoltaic and wind power, and promotes the upgrade of microgrid from passive stability maintenance to active precise control.

[0069] In another embodiment of the present application, a dynamic cooperative control system of gas turbine and microgrid is also provided, as shown in Figure 2 The system is applied to the foregoing method, and the system at least includes the following: The monitoring unit is configured to construct a microgrid real-time monitoring module, continuously collect real-time output of distributed energy, real-time demand of controllable load, microgrid bus voltage and frequency, gas turbine operation state parameters, and energy storage system charging and discharging state parameters in the microgrid. The distributed energy at least includes photovoltaic power supply, wind power supply and gas turbine, and the gas turbine at least includes a micro gas turbine adapted to a community / park microgrid or a small gas turbine adapted to an industrial park microgrid. The real-time monitoring module is equipped with a data preprocessing unit, which reduces noise of the collected high-frequency fluctuation data through a filtering algorithm to ensure that the data accuracy meets the subsequent control requirements. The control unit is configured to calculate a microgrid real-time power difference value based on the pre-processed data, the real-time power difference value being a difference between total output of the distributed energy and total demand of the controllable load; call a preset adaptive neuro-fuzzy inference system (ANFIS) to take the real-time power difference value, the microgrid bus voltage deviation value and the frequency deviation value as input variables, determine whether the real-time power difference value exceeds a preset power fluctuation threshold value and determine whether the microgrid bus voltage deviation value or the frequency deviation value exceeds a preset stability threshold value through a fuzzy rule base and a neural network training model of the ANFIS; if none of the real-time power difference value, the microgrid bus voltage deviation value and the frequency deviation value exceeds, maintain the current output of the gas turbine and the current charge and discharge state of the energy storage system; if any one of the real-time power difference value, the microgrid bus voltage deviation value and the frequency deviation value exceeds, trigger a dynamic cooperative control instruction.

[0070] The dynamic cooperative control system of the gas turbine and the microgrid can implement each step in the foregoing method embodiment, which will not be described here again.

[0071] In still another embodiment of the present application, an electronic device is also provided, as shown in Figure 3 The electronic device includes a processor 1110, a communication interface 1120, a memory 1130 and a communication bus 1140, wherein the processor 1110, the communication interface 1120 and the memory 1130 complete mutual communication through the communication bus; the memory 1130 is used to store a computer program; and the processor 1110 is used to execute the program stored on the memory 1130 to implement the dynamic cooperative control method of the gas turbine and the microgrid. The communication bus 1140 mentioned above can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc.

[0072] For ease of representation, Figure 3Only one thick line is used to represent a bus, but this is merely for the sake of clarity and only one bus or one type of bus can be present. The communication interface 1120 is configured to enable communication between the electronic device and other devices. The memory 1130 can include a random access memory (RAM) and can also include a non-volatile memory, such as at least one disk storage. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor. The processor 1110 described above can be a general-purpose processor, including an artificial intelligence processor, a graphics processing unit (GPU), a machine learning unit (MLU), a central processing unit (CPU), a network processor (NP), etc. It can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0073] Correspondingly, the embodiments of the present application further provide a computer readable storage medium storing a computer program, which, when executed, can implement each step that can be executed by an electronic device in the method embodiments described above.

Claims

1. A method for dynamic coordinated control of a microgrid and a combustion engine, characterized in that, The method comprises: The microgrid real-time monitoring module is constructed to continuously collect real-time output of distributed energy in the microgrid, real-time demand of controllable load, microgrid bus voltage and frequency, operating state parameters of the combustion engine, and state parameters of charge and discharge of the energy storage system; the distributed energy at least includes photovoltaic power supply, wind power supply and combustion engine, and the combustion engine at least includes a micro combustion engine adapted to the community / park microgrid or a small combustion engine adapted to the industrial park microgrid; the real-time monitoring module is provided with a data preprocessing unit, high-frequency fluctuation data collected are denoised by a filtering algorithm to ensure that data accuracy meets subsequent control requirements; The microgrid real-time power difference value is calculated based on the preprocessed data, and the real-time power difference value is the difference between the total output of the distributed energy and the total demand of the controllable load; The preset adaptive neuro-fuzzy inference system (ANFIS) is called to take the real-time power difference value, the microgrid bus voltage deviation value and the frequency deviation value as input variables, and to determine whether the real-time power difference value exceeds the preset power fluctuation threshold value and whether the microgrid bus voltage deviation value or the frequency deviation value exceeds the preset stability threshold value through the fuzzy rule base and the neural network training model of the ANFIS; If none of the real-time power difference value, the microgrid bus voltage deviation value and the frequency deviation value exceeds, the current output of the combustion engine and the current charge and discharge state of the energy storage system are maintained; if any one of the real-time power difference value, the microgrid bus voltage deviation value and the frequency deviation value exceeds, a dynamic cooperative control instruction is triggered.

2. The method according to claim 1, wherein After the dynamic cooperative control instruction is triggered, the following steps are further included: The energy storage system is preferentially controlled to perform millisecond / second response adjustment, the power adjustment direction and adjustment amount are determined based on the real-time power difference value; when the real-time power difference value is positive, it is determined that the microgrid is in an energy surplus state, and the energy storage system is controlled to increase the charging power to absorb the excess power; When the real-time power difference value is negative, it is determined that the microgrid is in an energy deficiency state, and the energy storage system is controlled to increase the discharging power to make up for the power gap; The model predictive control (MPC) algorithm is adopted to predict future short-term power fluctuation according to the change trend of the real-time power difference value, and dynamically correct the charge and discharge power instruction of the energy storage system to suppress short-term fluctuation caused by sudden drop of photovoltaic / wind power output or sudden increase of load.

3. The method according to claim 2, wherein The model predictive control (MPC) algorithm is adopted to predict future short-term power fluctuation according to the change trend of the real-time power difference value, and dynamically correct the charge and discharge power instruction of the energy storage system to suppress short-term fluctuation caused by sudden drop of photovoltaic / wind power output or sudden increase of load, which comprises: A prediction model of the MPC algorithm is constructed, and the real-time power difference value of the microgrid, the real-time output change rate of the photovoltaic power supply, the real-time output change rate of the wind power supply and the real-time demand change rate of the controllable load are taken as input variables of the model; the training samples of the prediction model are time series data of power difference value, photovoltaic / wind power fluctuation data and load demand fluctuation data in the historical operation of the microgrid, and the model parameters are optimized through offline training to make the prediction accuracy of the model meet the prediction demand of short-term power fluctuation. Based on the preprocessed data output by the microgrid real-time monitoring module, real-time power difference, photovoltaic power output rate, wind power output rate, and controllable load demand rate are obtained in real time and input into the trained MPC prediction model to output a microgrid power fluctuation prediction curve in the short term; According to the power fluctuation prediction curve, the theoretical charge / discharge power correction amount required by the energy storage system is calculated; if the prediction curve shows that the power difference will increase, the charge power correction amount or the discharge power correction amount of the energy storage system is increased; if the prediction curve shows that the power difference will tend to be flat, the charge power correction amount or the discharge power correction amount of the energy storage system is reduced; The charge / discharge power correction amount is superimposed on the initial charge / discharge power instruction of the energy storage system to generate a final energy storage system charge / discharge power control instruction, and the energy storage system is controlled to operate according to the instruction to suppress short-term fluctuations caused by sudden drops in photovoltaic / wind power output or sudden increases in load; During the operation of the MPC algorithm, the latest real-time power difference collected by the microgrid real-time monitoring module is continuously compared with the predicted value of the MPC prediction model to calculate the prediction deviation, and if the deviation exceeds the preset deviation threshold, the parameters of the MPC prediction model are corrected online.

4. The method according to claim 2, wherein After the dynamic collaborative control instruction is triggered, the following steps are further included: When the energy storage system starts to respond to the adjustment, a gas turbine output adjustment instruction is generated based on the real-time power difference and the current operation mode of the microgrid, combined with a long short-term memory network (LSTM) prediction model. The LSTM prediction model uses historical power difference data, photovoltaic / wind power output prediction values, and load prediction values as training samples to output a microgrid power demand prediction result in the medium and long term in real time; If the microgrid is in a grid-connected mode, the fuel supply amount of the gas turbine is adjusted to gradually adjust the gas turbine output, so that the gas turbine output and the energy supplement / absorption effect of the energy storage system form a time sequence collaboration, and the real-time power difference of the microgrid is stably controlled within a preset range, combined with the grid interaction power limit and the microgrid power demand prediction result; If the microgrid is in an off-grid mode, the starting timing or output increase range of the gas turbine is adjusted based on the remaining capacity of the energy storage system and the discharge duration predicted by the LSTM, with the microgrid bus voltage and frequency stability as the core target; The real-time monitoring module continuously tracks the changes in the microgrid bus voltage, frequency, and power difference. When it is monitored that the microgrid bus voltage and frequency continuously return to the preset stable threshold, and the real-time power difference continuously maintains within the preset power fluctuation threshold, a PID closed-loop adjustment algorithm is called to control the energy storage system to reduce the charge / discharge power to a normal charge / discharge state, and the gas turbine output is stably controlled at the current adjusted operating point, completing a dynamic collaborative control process.

5. The method according to claim 4, wherein The generation of the gas turbine output adjustment instruction based on the real-time power difference and the current operation mode of the microgrid, combined with the long short-term memory network (LSTM) prediction model, includes: The LSTM prediction model architecture with attention mechanism is constructed, which includes an input layer, an attention layer, a bidirectional LSTM hidden layer, a regularization layer, and an output layer; wherein the input layer receives the microgrid historical power difference data, photovoltaic power output prediction value, wind power output prediction value, and controllable load demand prediction value; the attention layer strengthens the feature expression of key variables by calculating the contribution weight of each input variable to the microgrid power demand; the bidirectional LSTM hidden layer is used to capture the bidirectional dependency of the input time series data, and the regularization layer suppresses model overfitting through an adaptive weight decay mechanism; The LSTM prediction model is trained in a distributed manner using a federated learning framework, with historical operation data of multiple microgrids of the same type as distributed training samples. The model gradient of each microgrid is aggregated through a federated averaging algorithm. During the training process, the mean absolute percentage error (MAPE) is used as the loss function, and the training step size is dynamically adjusted through an adaptive learning rate optimization algorithm until the model prediction accuracy meets the preset requirements. The model input data, including historical power difference time series data, photovoltaic / wind power output prediction value, and controllable load demand prediction value, are obtained through the microgrid real-time monitoring module. After standardization preprocessing, the input data are input into the trained LSTM prediction model. The LSTM prediction model dynamically allocates variable weights through the attention layer, extracts time series features through the bidirectional LSTM hidden layer, and finally outputs the microgrid power demand prediction results for the future medium and long term. The power demand prediction results are dynamically corrected in combination with the current operation mode and real-time power difference of the microgrid. In grid-connected mode, the large grid interactive power constraint factor is introduced, and the prediction results are adjusted according to the grid dispatching instruction to ensure that the gas turbine output adjustment meets the grid interactive power limit value. In off-grid mode, the energy storage system state constraint factor is introduced, and the prediction results are corrected based on the remaining capacity and charge / discharge characteristics of the energy storage system to avoid time mismatch between gas turbine output and energy storage compensation. Based on the corrected power demand prediction results, the gas turbine output adjustment instruction is generated in combination with the gas turbine's own operation characteristics. The instruction includes the gas turbine target output reference value, output adjustment gradient, and start / stop timing logic.

6. The method according to claim 5, wherein The microgrid bus voltage, frequency, and power difference are continuously tracked through the real-time monitoring module. When the microgrid bus voltage and frequency continuously return to the preset stable threshold, and the real-time power difference continuously maintains within the preset power fluctuation threshold, the PID closed-loop adjustment algorithm is called to control the energy storage system to reduce the charge / discharge power to the normal charge / discharge state, including: The fuzzy PID regulation model fused with reinforcement learning comprises a fuzzy reasoning unit and a reinforcement learning optimization unit; the fuzzy reasoning unit takes the voltage deviation and rate of change of the voltage deviation, the frequency deviation and rate of change of the frequency deviation, and the real-time power difference deviation and rate of change of the real-time power difference as inputs, and maps the input variables to the adjustment interval of the PID parameters through a preset fuzzy rule base; the reinforcement learning optimization unit takes the smooth transition of the charge and discharge power of the energy storage system to the normal state and the absence of overshoot fluctuation of the microgrid parameters as a reward function, dynamically optimizes the PID parameter adjustment interval output by the fuzzy reasoning unit through real-time interaction with the microgrid operating environment, and improves the parameter adaptability; The real-time data of the microgrid bus voltage, frequency and real-time power difference are obtained through the microgrid real-time monitoring module, the deviations and rates of change of the parameters from the preset thresholds are calculated, and the deviations and rates of change are input into the fuzzy PID regulation model fused with reinforcement learning; the model outputs the optimal PID parameters under the current working condition after fuzzy reasoning and reinforcement learning optimization; Based on the output PID parameters, a charge and discharge power regulation closed loop of the energy storage system is constructed; the difference between the current charge and discharge power of the energy storage system and the normal charge and discharge power is taken as the PID regulation deviation, the charge and discharge power regulation amount of the energy storage system is calculated through the PID algorithm, and the power control instruction of the energy storage system is generated; at the same time, a microgrid parameter feedback mechanism is introduced, and if the microgrid bus voltage, frequency or real-time power difference deviates from the preset threshold during the regulation process, the PID regulation deviation is corrected in real time to avoid the regulation process causing fluctuations in the microgrid parameters; The control energy storage system gradually reduces the charge and discharge power according to the power control instruction, and synchronously tracks the change of the charge and discharge power of the energy storage system and the stability of the microgrid parameters through the microgrid real-time monitoring module; if the charge and discharge power of the energy storage system approaches the normal charge and discharge power range, the PID parameter fine-tuning mode is triggered, the regulation rate is reduced by reducing the proportional coefficient and the integral coefficient, and the smooth transition of the energy storage system power to the normal state is ensured; When the microgrid real-time monitoring module monitors that the charge and discharge power of the energy storage system continuously and stably stays within the normal charge and discharge power range, and the microgrid bus voltage, frequency and real-time power difference all continuously maintain within the preset threshold, it is determined that the regulation of the energy storage system is completed, the PID closed-loop regulation algorithm is stopped, and the current energy storage system operating state is maintained, thereby providing a basic condition for the next dynamic collaborative control.

7. The method of dynamic co-ordinated control of a microgrid and a combustion engine according to claim 1, characterized in that, The microgrid real-time power difference is calculated based on the preprocessed data, and the real-time power difference is the difference between the total output of the distributed energy and the total demand of the controllable load, which comprises: A distributed energy total output calculation model is constructed, which introduces a wavelet transform-Kalman filter fusion algorithm to perform secondary noise reduction and trend extraction on the preprocessed distributed energy real-time output data output by the microgrid real-time monitoring module; the wavelet transform unit separates the high-frequency noise and low-frequency effective signal in the output data through multi-scale decomposition, and the Kalman filter unit performs dynamic filtering on the decomposed low-frequency effective signal to eliminate fluctuations caused by data acquisition delay and environmental interference, and outputs the smooth output value of each distributed energy; A controllable load total demand calculation model is constructed, and an adaptive weight distribution algorithm is used to aggregate the real-time demand data of each controllable load after preprocessing; according to the importance level and operating characteristics of each controllable load, the calculation weight of the demand data is dynamically distributed, and a transient value correction factor is additionally superimposed on the impact load demand data, so as to avoid the calculation deviation of the total demand caused by the sudden demand of a single load, and to output the smooth value of the controllable load total demand; A power data time sequence alignment mechanism is established, and through the time synchronization unit of the microgrid real-time monitoring module, the timestamp of the smooth output value of each distributed energy and the smooth value of the controllable load total demand is calibrated, so as to ensure that all the power data participating in the calculation correspond to the same time dimension, and to eliminate the calculation error caused by the asynchronous data collection; The real-time power difference value is calculated, the smooth output value of each distributed energy is summed to obtain the total output of the distributed energy, and then the total output of the distributed energy is subtracted from the smooth value of the controllable load total demand to obtain the real-time power difference value; A power difference dynamic correction mechanism is introduced, a power difference correction model is constructed based on the historical operation data of the microgrid, the current microgrid operation mode and the distributed energy output fluctuation coefficient are taken as input variables, a radial basis function neural network RBFNN is used to output a power difference correction coefficient, the power difference correction coefficient is multiplied by the real-time power difference value to obtain the real-time power difference value of the microgrid.

8. The method according to claim 7, wherein According to the importance level and operating characteristics of each controllable load, the calculation weight of the demand data is dynamically distributed, and a transient value correction factor is additionally superimposed on the impact load demand data, so as to avoid the calculation deviation of the total demand caused by the sudden demand of a single load, and to output the smooth value of the controllable load total demand, including: A controllable load classification and feature extraction model is constructed, a variational mode decomposition VMD algorithm is used to decompose the real-time demand data of each controllable load after preprocessing, and the stationary component and the impact component in the load demand are separated; the importance coefficient and the impact characteristic coefficient of each load are extracted in combination with the preset load importance level of the microgrid load management system, and a load feature matrix is formed; A dynamic weight distribution model is constructed based on a deep Q network DQN, the load feature matrix, the current operation mode of the microgrid and the real-time output stability of the distributed energy are taken as state inputs, and the minimization of the total demand calculation deviation of the controllable load is taken as a reward function, the mapping relationship between the load features and the calculation weight is established through the interaction training of the DQN and the microgrid operation environment; in the training process, the weight update strategy is optimized through the experience replay mechanism, so as to ensure that the adaptive load demand calculation weight can be output under different working conditions; For the impact load demand data, a transient value correction model is constructed; the peak time, duration and decay characteristics of the impact load are extracted through Hilbert-Huang transform HHT for the impact component obtained by VMD decomposition; based on the historical impact load response data of the microgrid, an impact load transient value correction factor library is established, the peak intensity and duration of the impact component are taken as indexes, and the corresponding correction factor is matched, the correction factor dynamically changes with the decay trend of the impact load, and over-correction or under-correction caused by a single fixed correction value is avoided. The weighted demand value of each controllable load is calculated, the calculated weight is multiplied by the real-time demand data of the corresponding load, and a transient value correction factor is additionally superimposed on the impact load to obtain the weighted demand value of each controllable load; A controllable load total demand smoothing model is constructed, a sliding window filtering algorithm is used to aggregate and smooth the weighted demand values of each controllable load, and the window length is dynamically adjusted according to the microgrid load fluctuation frequency; a load demand trend prediction mechanism is introduced, the load demand trend in a short time in the future is predicted through a gated recurrent unit (GRU), and the aggregated total demand value is trend corrected to output a controllable load total demand smoothing value.

9. The method of dynamic co-ordinated control of a microgrid and a combustion engine according to claim 1, characterized in that, The preset adaptive neuro-fuzzy inference system (ANFIS) is called to take the real-time power difference value, the microgrid bus voltage deviation value and the frequency deviation value as input variables, and to determine whether the real-time power difference value exceeds the preset power fluctuation threshold value and whether the microgrid bus voltage deviation value or the frequency deviation value exceeds the preset stability threshold value through the fuzzy rule base and the neural network training model of the ANFIS, including: The real-time power difference value, the bus voltage deviation value and the frequency deviation value data in the historical operation of the microgrid are clustered and analyzed by using a fuzzy C-means clustering (FCM) algorithm to generate initial fuzzy subsets and membership functions of the ANFIS fuzzy rule base; The threshold-exceeding samples and normal samples in the historical data are taken as a training set to train the ANFIS, the membership function parameters and the neural network connection weights are optimized by using a back propagation algorithm, and the identification accuracy of the model to the threshold-exceeding is made to meet preset requirements; The trained ANFIS is called, and the real-time calculated microgrid real-time power difference value, bus voltage deviation value and frequency deviation value are input into the model; fuzzy inference is performed through the fuzzy rule base to output fuzzy judgment results corresponding to each input variable; The fuzzy judgment results are de-fuzzified by the ANFIS to be converted into clear numerical judgment conclusions to determine whether the real-time power difference value exceeds the preset power fluctuation threshold value and whether the bus voltage deviation value or the frequency deviation value exceeds the preset stability threshold value.

10. A dynamic co-ordinated control system of a microgrid and a combustion engine, characterized in that, The system is applied to the dynamic collaborative control method of the gas turbine and the microgrid according to any one of claims 1 to 9, and the system comprises: A monitoring unit is configured to construct a microgrid real-time monitoring module, continuously collect distributed energy real-time output, controllable load real-time demand, microgrid bus voltage and frequency, gas turbine operating state parameters and energy storage system charging and discharging state parameters in the microgrid; the distributed energy at least includes photovoltaic power, wind power and a gas turbine, and the gas turbine at least includes a micro gas turbine adapted to a community / park microgrid or a small gas turbine adapted to an industrial park microgrid; the real-time monitoring module is equipped with a data preprocessing unit, high-frequency fluctuation data collected by the unit is denoised through a filtering algorithm to ensure that the data accuracy meets the subsequent control requirements; The control unit is configured to calculate a micro-grid real-time power difference value based on the pre-processed data, the real-time power difference value being a difference between total output of the distributed energy and total demand of the controllable load; call a preset adaptive neuro-fuzzy inference system (ANFIS) to take the real-time power difference value, the micro-grid bus voltage deviation value and the frequency deviation value as input variables, judge whether the real-time power difference value exceeds a preset power fluctuation threshold value and whether the micro-grid bus voltage deviation value or the frequency deviation value exceeds a preset stability threshold value through a fuzzy rule base and a neural network training model of the ANFIS; if none of the real-time power difference value, the micro-grid bus voltage deviation value and the frequency deviation value exceeds, maintain the current output of the gas turbine and the current charging and discharging state of the energy storage system; if any one of the real-time power difference value, the micro-grid bus voltage deviation value and the frequency deviation value exceeds, trigger a dynamic cooperative control instruction.

Images