Intelligent electric power comprehensive operation management system
By constructing a hybrid prediction model and a multi-objective optimization module, the problems of photovoltaic power output prediction accuracy and optimization strategy were solved, achieving efficient photovoltaic consumption and stable grid operation.
Patent Information
- Application Number
- CN202511095294.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-06
- Publication Date
- 2025-11-07
AI Technical Summary
Existing photovoltaic power output prediction models are too simplistic or lack physical constraints, resulting in insufficient prediction accuracy, imperfect optimization strategies, difficulty in simultaneously considering curtailment rate, grid losses, and energy storage costs, and a lack of real-time dynamic adjustment mechanisms.
A physical mechanism model equivalent to a single diode and a spatiotemporal attention mechanism LSTM deep learning model are constructed. The weights are dynamically allocated by a Bayesian probability model and a third-generation non-dominated sorting genetic algorithm is used to generate a Pareto optimal solution set. The short-term photovoltaic power output prediction curve and optimization strategy are output.
It improves the accuracy of photovoltaic output forecasting, balances curtailment rate, grid losses and energy storage costs, dynamically responds to weather fluctuations, and ensures stable grid operation.
Smart Images
Figure CN120914795A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of power integrated operation management, and particularly relates to a smart power integrated operation management system. BACKGROUND
[0002] With the rapid development of the new energy industry, the proportion of new energy such as photovoltaic in the power system is gradually increasing. However, photovoltaic has strong intermittency, which leads to large power fluctuation of the power grid, and the traditional dispatching mode is difficult to accurately match supply and demand, seriously affecting the stable operation of the power grid and the efficient consumption of new energy. At present, there are some researches and applications for photovoltaic output prediction and consumption optimization, but there are still the following problems: the prediction model is single, either only relying on a physical mechanism model, which is difficult to cope with nonlinear changes under complex weather conditions, or only using a deep learning model, which lacks physical constraints and the prediction accuracy needs to be improved. In terms of optimization strategy, the application of multi-objective optimization algorithm is not perfect, it is difficult to consider multiple objectives such as light rejection rate, power grid loss and energy storage cost at the same time, and lacks a real-time dynamic adjustment mechanism, when there is a large deviation between the actual situation and the prediction, it cannot respond in time. Therefore, there is an urgent need for a smart power integrated operation management system to solve the above problems.
[0003] To this end, the application provides a smart power integrated operation management system. SUMMARY
[0004] In order to make up for the deficiencies of the prior art and solve at least one technical problem proposed in the background art.
[0005] The technical scheme adopted by the application to solve its technical problems is: In a first aspect, the application provides a smart power integrated operation management system, comprising: a data acquisition module configured to obtain historical operation data of a photovoltaic power station; a hybrid prediction module configured to construct a single diode equivalent physical mechanism model and a spatiotemporal attention mechanism LSTM deep learning model, combine the historical operation data of the photovoltaic power station, dynamically allocate the weights of the physical mechanism model and the deep learning model through a Bayesian probability model, and output a short-term photovoltaic output prediction curve; a multi-objective optimization module configured to construct an optimization target model, constrain the target optimization model, generate a Pareto optimal solution set using a third-generation non-dominated sorting genetic algorithm, and output a short-term distributed power output plan, an energy storage charge-discharge curve, and a flexible load adjustment instruction.
[0006] As one of the embodiments, the historical operation data of the photovoltaic power station specifically comprises: photovoltaic power station parameter data and photovoltaic power station environment data; The photovoltaic power station parameter data includes hourly power generation, component temperature, inverter efficiency, photovoltaic output value, and the photovoltaic power station environment data includes irradiance, cloud cover, wind speed, relative humidity and ambient temperature.
[0007] As one of the embodiments, the physical mechanism model is specifically: The physical mechanism model is specifically: , wherein, is the photo-generated current, is the photovoltaic power station output voltage, is the reverse saturation current, is the equivalent series resistance of the battery in the photovoltaic power station, is the output current of the photovoltaic power station, is the ideality factor reflecting the non-ideality of the PN junction, is the Boltzmann constant, is the absolute temperature, is the electronic charge, is the equivalent parallel resistance of the battery in the photovoltaic power station.
[0008] As one of the embodiments, the physical mechanism model functions as: According to the irradiance and ambient temperature collected in real time, the irradiance and ambient temperature are divided into different types of working conditions, for each type of working condition, the physical mechanism model parameters are calculated and stored, the corresponding working condition is matched according to the irradiance and ambient temperature collected in real time, and the irradiance and ambient temperature are substituted into the physical mechanism model to output the photovoltaic output theoretical value.
[0009] As one of the embodiments, the deep learning model is specifically: In the historical operation data, 5-dimensional data including photovoltaic output value, irradiance, ambient temperature, relative humidity and wind speed are selected, X-minute time steps are set, 4C time step data are constructed as input sequences, and the photovoltaic output value of the future C time steps is predicted; Setting the number of cores to 64, the embedding layer output is 64x64, the attention weight of the feature dimension is calculated through 1x1 convolution, the spatial attention output is consistent with the embedding layer output layer, a single-layer LSTM is used, 64 units are set for each layer to extract the time dependence, the attention weight of the time step dimension is calculated, and the attention weight of the feature dimension and the attention weight of the time step dimension are fused by layer-by-layer multiplication. The optimal fusion model of the deep learning model is determined through the validation set, and the photovoltaic output value of the future C time steps is predicted according to the optimal fusion model.
[0010] As one of the embodiments, the specific process of dynamically allocating the weights of the physical mechanism model and the deep learning model through the Bayesian probability model is: According to historical operation data of the photovoltaic power station, a fluctuation amplitude of irradiance in the past N time steps is calculated, and the fluctuation amplitude of irradiance is specifically an irradiance change rate; a fluctuation amplitude of ambient temperature in the past N time steps is calculated, and the fluctuation amplitude of ambient temperature is specifically an ambient temperature change amount; When the irradiance change rate is less than an irradiance change rate threshold value, and the ambient temperature change amount is less than an ambient temperature change amount threshold value, it is determined that the power station working condition is a meteorological stable working condition; When the irradiance change rate is greater than the irradiance change rate threshold value, and the ambient temperature change amount is greater than the ambient temperature change amount threshold value, it is determined that the power station working condition is a meteorological unstable working condition; In the meteorological stable working condition, the initial weight of the deep learning model is set as , the initial weight of the physical mechanism model is set as , and the sum of the initial weight of the deep learning model and the initial weight of the physical mechanism model is 1; In the meteorological unstable working condition, the initial weight of the deep learning model is set as , the initial weight of the physical mechanism model is set as , and the sum of the initial weight of the deep learning model and the initial weight of the physical mechanism model is 1; In the initial weight range in the meteorological stable working condition and the meteorological unstable working condition, the fusion error is calculated through Bayesian calculation, the deep learning model prediction error and the physical mechanism model prediction error in the historical operation data are respectively counted in the meteorological stable working condition and the meteorological unstable working condition, and the best fusion weight combination is output according to the deep learning model prediction error, the physical mechanism model prediction error and the fusion error.
[0011] As one of the embodiments, the specific process of outputting the short-term photovoltaic output prediction curve is as follows: Based on the best fusion weight combination, the physical mechanism model and the deep learning model are fused to output the short-term photovoltaic output prediction curve.
[0012] As one of the embodiments, the specific process of constructing the optimization target model is as follows: The optimization target model includes: a light rejection rate model, a power grid loss model, and a storage charging and discharging cost model. The light rejection rate model is as follows: , wherein is a photovoltaic output theoretical value, is an actual grid-connected power, is a time step, is a time step index; The power grid loss model is as follows: , wherein is the total number of lines, is a line current, is a line resistance, is a line total number index; energy storage charging and discharging cost model: wherein, is an energy storage charging power, is an energy storage discharging power, , is a time-of-use electricity price, and 0.25 is a time step, is a time step, is a time step index.
[0013] As one of the embodiments, the specific process of generating a Pareto optimal solution set by using the third generation non-dominated sorting genetic algorithm is: Based on the target optimization model, the total number of time steps of the light rejection rate model, the grid loss model, and the energy storage charging and discharging cost model in the target optimization model is obtained, and the population size is set according to the total number of time steps; The non-dominated sorting is used to filter the front solution set which is not better than itself, and the crowding degree of the solutions in the front is calculated to retain diversity; Through simulated binary crossover and polynomial mutation simulation evolution, offspring solutions are generated, and then, taking the reference points of the light rejection rate tending to 0, the grid loss tending to 0, and the cost tending to 0, solutions with close distance to the reference points and high crowding degree are selected to form the next generation, and the iteration is repeated until the target value change rate of the optimal solution of the continuous 5 generations is <1%, and finally, the Pareto optimal solution set covering the multi-objective trade-off relationship is output.
[0014] As one of the embodiments, the specific process of outputting the short-term distributed power output plan, the energy storage charging and discharging curve, and the flexible load adjustment instruction is: Based on the generated Pareto optimal solution set, solutions meeting the preset conditions are filtered from the Pareto optimal solution set, and the short-term distributed power output plan, the energy storage charging and discharging curve, and the flexible load adjustment instruction are output.
[0015] In a second aspect, the present application provides a smart power comprehensive operation management method, comprising: S1: obtaining historical operation data of a photovoltaic power station; S2: constructing a single diode equivalent physical mechanism model and a spatiotemporal attention mechanism LSTM deep learning model, combining the historical operation data of the photovoltaic power station, dynamically distributing the physical mechanism model and the deep learning model weight through the Bayesian probability model, and outputting a short-term photovoltaic output prediction curve; S3: constructing an optimization target model, and constraining the target optimization model, generating a Pareto optimal solution set by using the third generation non-dominated sorting genetic algorithm, and outputting a short-term distributed power output plan, an energy storage charging and discharging curve, and a flexible load adjustment instruction.
[0016] The beneficial effects of the present application are as follows: 1. Dynamically allocate the weights of the physical mechanism model and the deep learning model through the Bayesian probability model, adapt to the stable and unstable working conditions of the weather, fuse the advantages of both, and output more accurate short-term photovoltaic output prediction curves to provide reliable basis for subsequent optimization.
[0017] 2. Construct a multi-objective model of light rejection rate, power grid loss, and energy storage charging and discharging cost, generate a Pareto optimal solution set combining the third generation non-dominated sorting genetic algorithm, balance the conflicts of each target, effectively reduce the light rejection rate, reduce the power grid loss, control the energy storage cost, and improve the photovoltaic consumption efficiency and system economy.
[0018] 3. Divide the optimization period by time steps, match the power dispatch rhythm, and ensure that the strategy meets the requirements of power grid safety and equipment capacity. The optimization result can be directly converted into dispatching instructions, which has strong landing performance, dynamically responds to meteorological fluctuations and actual operation deviations, adjusts the strategy through rolling optimization, improves the system's response to photovoltaic intermittency, and ensures the stable operation of the power grid. BRIEF DESCRIPTION OF DRAWINGS
[0019] The present application will be further described below with reference to the accompanying drawings.
[0020] Figure 1 is a system module diagram of a smart power comprehensive operation management system of the present application; Figure 2 is a step flow chart of a smart power comprehensive operation management method of the present application. DETAILED DESCRIPTION
[0021] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application will be further described below with reference to the specific embodiments.
[0022] Embodiment 1 As shown in the drawings, the smart power comprehensive operation management system according to the embodiment of the present application comprises: Figure 1 A data acquisition module acquires historical operation data of a photovoltaic power station. The data acquisition module acquires historical operation data of a photovoltaic power station. In the data acquisition module, the first specific, the historical operation data of the photovoltaic power station specifically includes: photovoltaic power station parameter data and photovoltaic power station environment data. The photovoltaic power station parameter data includes the parameter data of the photovoltaic power station itself, including hourly power generation, component temperature, inverter efficiency, and photovoltaic output value; and the photovoltaic power station environment data includes irradiance, cloud cover, wind speed, relative humidity, and ambient temperature. Exemplary, the SCADA (Supervisory Control And Data Acquisition) and EMS (Energy Management System) of the photovoltaic power station will store the historical operation data of the photovoltaic power station for a long time, including hourly power generation, component temperature, inverter efficiency, photovoltaic output value, irradiance, cloud cover, wind speed, relative humidity, ambient temperature, through the database of the system background, the historical data is exported according to the time interval, or the API interface provided by the system is used to batch acquire; Through the special meteorological station deployed in the photovoltaic power station, the environmental data of the photovoltaic power station are obtained, and the irradiance, cloud cover, wind speed, relative humidity, and ambient temperature are uploaded; In addition, the SCADA (Supervisory Control And Data Acquisition) and EMS (Energy Management System) of the photovoltaic power station can access synchronous real-time load data, current energy storage SOC, and dynamic parameters such as time-of-use electricity price. It can also load static data such as power grid topology, line parameters, and flexible load response characteristics from the photovoltaic power station system configuration module. The SCADA (Supervisory Control And Data Acquisition) and EMS (Energy Management System) of the photovoltaic power station, and the special meteorological station of the photovoltaic power station can access the real-time data information collected by each other through the industrial bus.
[0023] Hybrid prediction module: construct a single diode equivalent physical mechanism model and a spatiotemporal attention mechanism LSTM deep learning model, combine the historical operation data of the photovoltaic power station, dynamically allocate the weights of the physical mechanism model and the deep learning model through the Bayesian probability model, and output the short-term photovoltaic output prediction curve; In the hybrid prediction module, first, the physical mechanism model is specifically: Using the electrical characteristics of the photovoltaic cell, a single diode equivalent circuit model is used to simulate the electrical characteristics of the photovoltaic cell, and a physical mechanism model is constructed. The physical mechanism model is specifically: , wherein, is the photo-generated current, which is calculated by irradiance and ambient temperature, is the output voltage of the photovoltaic power station, is the reverse saturation current, which is the leakage current of the PN junction under no light, and is related to the ambient temperature, is the equivalent series resistance of the battery in the photovoltaic power station, is the output current of the photovoltaic power station, is the ideality factor, which reflects the degree of non-ideality of the PN junction, is the Boltzmann constant, is the absolute temperature, is the electronic charge, is the equivalent parallel resistance of the battery in the photovoltaic power station; According to the irradiance and the environmental temperature collected in real time, the irradiance and the environmental temperature are divided into different types of working conditions, for each type of working condition, the physical mechanism model parameters are calculated and stored, the corresponding working condition is matched according to the irradiance and the environmental temperature collected in real time, and the irradiance and the environmental temperature are substituted into the physical mechanism model, and the theoretical value of the photovoltaic output is output; For example, according to the irradiance , the irradiance is divided into three types of irradiance intensity, i.e., low irradiance intensity , medium irradiance intensity , and high irradiance intensity , and the environmental temperature is divided into three types of environmental temperature, i.e., low environmental temperature , medium environmental temperature , and high environmental temperature ; The irradiance is divided into three types of irradiance intensity, i.e., low irradiance intensity, medium irradiance intensity, and high irradiance intensity, and the environmental temperature is divided into three types of environmental temperature, i.e., low environmental temperature, medium environmental temperature, and high environmental temperature, and the nine types of working conditions are obtained by permutation and combination, for example, the low environmental temperature and the low irradiance intensity are one type of working condition, the physical mechanism model parameters are calculated and stored, the corresponding working condition is matched according to the irradiance and the environmental temperature collected in real time, and the irradiance and the environmental temperature are substituted into the physical mechanism model, and the theoretical value of the photovoltaic output is output; It should be noted that the different types of working conditions are divided, and the present application does not limit this, which can be divided according to the specific situation; In the hybrid prediction module, the second specific deep learning model is specifically: In the historical operation data, 5-dimensional data including photovoltaic output value, irradiance, environmental temperature, relative humidity, and wind speed are selected, the time step is set to 15 minutes, 64 time step data are constructed as input sequences, and the photovoltaic output value of the future 16 time steps is predicted; The number of kernels is set to 64, the embedding layer output is 64x64, the attention weight of the feature dimension is calculated through 1x1 convolution, the spatial attention output is consistent with the embedding layer output layer, a single-layer LSTM is used, 64 units are set for each layer, the time dependence is extracted, the attention weight of the time step dimension is calculated, and the attention weight of the feature dimension and the attention weight of the time step dimension are fused by layer-by-layer multiplication; The optimal fusion model of the deep learning model is determined through the validation set, and the photovoltaic output value of the future 16 time steps is predicted according to the optimal fusion model; In the hybrid prediction module, the third specific process of dynamically allocating the weights of the physical mechanism model and the deep learning model through the Bayesian probability model is specifically: The photovoltaic output value is determined by irradiance and ambient temperature, the irradiance directly determines the photo-generated current, the ambient temperature affects the cell internal resistance and the reverse saturation current, by determining the fluctuation of irradiance and ambient temperature, determining whether the physical mechanism model is stable, the deep learning model fits the data rule, the error is smaller but depends on historical operation data; Specifically, according to the historical operation data of the photovoltaic power station, the fluctuation amplitude of the irradiance in the past 4 time steps (a time step is 15 minutes, aligned with the time step of the deep learning model) is calculated, and the fluctuation amplitude of the irradiance is specifically: the irradiance change rate; Further, the fluctuation amplitude of the ambient temperature in the past 4 time steps (a time step is 15 minutes, aligned with the time step of the deep learning model) is calculated, and the fluctuation amplitude of the ambient temperature is specifically: the ambient temperature change amount; The working condition of the power station is divided into stable working condition and sudden change working condition, specifically, when the irradiance change rate is less than the irradiance change rate threshold, and the ambient temperature change amount is less than the ambient temperature change amount threshold, the working condition of the power station is determined as the meteorological stable working condition; When the irradiance change rate is greater than the irradiance change rate threshold, and the ambient temperature change amount is greater than the ambient temperature change amount threshold, the working condition of the power station is determined as the meteorological unstable working condition; It should be noted that the irradiance change rate threshold and the ambient temperature change amount threshold are a reference value set by the technical personnel in the industry, and the setting of the irradiance change rate threshold and the ambient temperature change amount threshold is determined by the actual situation, and the present application does not limit it, and the setting of the irradiance change rate threshold and the ambient temperature change amount threshold should be able to filter out some extremely rare meteorological phenomena; Specific to the present application, the extremely rare meteorological phenomenon is two cases, specifically as follows: The irradiance change rate is greater than the irradiance change rate threshold, and the ambient temperature change amount is less than the ambient temperature change amount threshold, which violates the physical common sense; The irradiance change rate is less than the irradiance change rate threshold, and the ambient temperature change amount is greater than the ambient temperature change amount threshold, which also violates the physical common sense; In the meteorological stable working condition, the initial weight of the deep learning model is set as , ; the initial weight of the physical mechanism model is , ; that is, ; the sum of the initial weight of the deep learning model and the initial weight of the physical mechanism model is 1; In the meteorological unstable working condition, the initial weight of the deep learning model is set as , ; the initial weight of the physical mechanism model is , ; i.e. ; the sum of the initial weights of the deep learning model and the initial weights of the physical mechanism model is 1; The initial weights of the deep learning model and the initial weights of the physical mechanism model are not limited in the present application, and can be set according to actual conditions; In the initial weight range under the meteorological stable condition and the meteorological unstable condition, the fusion error is calculated by Bayes, the deep learning model prediction error and the physical mechanism model prediction error in the historical operation data under the meteorological stable condition and the meteorological unstable condition are respectively counted, the weights of the physical mechanism model and the deep learning model are dynamically allocated according to the deep learning model prediction error, the physical mechanism model prediction error and the fusion error, and the best fusion weight combination is output; In the hybrid prediction module, the fourth specific process of outputting the short-term photovoltaic output prediction curve is: Based on the best fusion weight combination, the physical mechanism model and the deep learning model are fused to output the short-term (16 time steps, one time step being 15 minutes) photovoltaic output prediction curve.
[0024] Multi-objective optimization module: construct an optimization target model, constrain the target optimization model, generate a Pareto optimal solution set by using the third generation non-dominated sorting genetic algorithm, and output a short-term distributed power output plan, a storage charging and discharging curve and a flexible load adjustment instruction; In the multi-objective optimization module, the first specific process of constructing the optimization target model is: The optimization target model includes: a light rejection rate model, a power grid loss model and a storage charging and discharging cost model; The light rejection rate model is: , wherein, is a photovoltaic output theoretical value, is an actual grid-connected power, is a time step, one step being 15 minutes, is a time step index; The power grid loss model is: , wherein, is the total number of lines, is a line current, is a line resistance, is a line total number index; The storage charging and discharging cost model is: , wherein, is a storage charging power, is a storage discharging power, , is a time-of-use electricity price, 0.25 is a time step length of 15 minutes, is a time step, is a time step index.
[0025] In the multi-objective optimization module, the second specific process for constraining the objective optimization model is as follows: The voltage of all nodes must be between 95% and 105% of the rated voltage, and the current of each line must not exceed 90% of the rated capacity. If the voltage of any node exceeds 95% to 105% of the rated voltage or the current of any line exceeds 90% of the rated capacity, the penalty mechanism will be triggered. Charging and discharging power must not exceed the rated power of the energy storage itself, and the state of charge (SOC) must be maintained between 20% and 80%. If any item fails to meet the requirements, a penalty mechanism will be triggered. The change in load power must not exceed the maximum adjustment capacity of the equipment; otherwise, a penalty mechanism will be triggered. It should be noted that the core function of the penalty mechanism is to allow non-compliant optimization solutions to be naturally eliminated in the algorithm iteration, and to guide the algorithm to prioritize compliant strategies through quantitative deduction. This invention does not limit the penalty mechanism and quantitative deduction. In the multi-objective optimization module, the specific process of generating the Pareto optimal solution set using the third-generation non-dominated sorting genetic algorithm is as follows: The conflict between multiple objectives is balanced through four steps: population initialization, non-dominated sorting, genetic operations, and environmental selection. Specifically, based on the target optimization model, the total number of time steps for the curtailment rate model, grid loss model, and energy storage charging and discharging cost model in the target optimization model is obtained, and the population size is set according to the total number of time steps. Next, the frontier solution set with no solution being superior to itself is filtered out by non-dominated sorting, and the crowding degree of the solutions within the frontier is calculated to preserve diversity. Then, through simulated binary crossover and polynomial mutation, offspring schemes are generated. Then, with the light abandonment rate approaching 0, network loss approaching 0, and cost approaching 0 as reference points, schemes that are close to the reference point and have high congestion are selected to form the next generation. This process is repeated until the target value change rate of the optimal scheme for 5 consecutive generations is <1%. Finally, a Pareto optimal solution set covering the multi-objective trade-off relationship is output. In the multi-objective optimization module, the third specific process for outputting short-term distributed power generation plans, energy storage charging and discharging curves, and flexible load adjustment instructions is as follows: Based on the generated Pareto optimal solution set, solutions that meet preset conditions are selected from the Pareto optimal solution set (such as the light rejection rate model). Power grid loss model Energy storage costs , i.e., the light rejection rate, the grid loss, and the energy storage charging and discharging cost are all less than the corresponding preset conditions, and outputting a short-term distributed power output plan: future grid-connected power of the photovoltaic power station (obtained through the above constraints and screening based on the light rejection rate model), an energy storage charging and discharging curve: a time sequence of energy storage charging and discharging power (obtained through the above constraints and screening based on the grid loss model), and a flexible load adjustment instruction: a power adjustment amount of the industrial load at each time step (obtained through the above constraints and screening based on the energy storage charging and discharging cost model); It should be noted that the preset conditions are not limited in the present application, and are set according to specific conditions. In some embodiments, the light rejection rate model , the grid loss model , and the energy storage cost ; In the present application, the length and number of time steps are not specifically limited, and can be flexibly set by those skilled in the art according to specific conditions.
[0026] Embodiment 2 As shown in Figure 2 , based on embodiment 1, the present application provides a smart power comprehensive operation management method, comprising: S1: obtaining historical operation data of a photovoltaic power station; The historical operation data of the photovoltaic power station specifically comprises: photovoltaic power station parameter data and photovoltaic power station environment data; The photovoltaic power station parameter data comprises hourly power generation, component temperature, inverter efficiency, and photovoltaic output value, and the photovoltaic power station environment data comprises irradiance, cloud cover, wind speed, relative humidity, and ambient temperature.
[0027] S2: constructing a single-diode equivalent physical mechanism model and a spatiotemporal attention mechanism LSTM deep learning model, combining the historical operation data of the photovoltaic power station, dynamically allocating the physical mechanism model and the deep learning model weight through a Bayesian probability model, and outputting a short-term photovoltaic output prediction curve; The physical mechanism model specifically comprises: The physical mechanism model specifically comprises: , wherein is the photo-generated current, is the output voltage of the photovoltaic power station, is the reverse saturation current, is the equivalent series resistance of the battery in the photovoltaic power station, is the output current of the photovoltaic power station, is the ideality factor, reflecting the non-ideal degree of the PN junction, is the Boltzmann constant, is the absolute temperature, For the electronic charge, For the equivalent parallel resistance of the cells in the photovoltaic power station.
[0028] The role of the physical mechanism model is: According to the irradiance and ambient temperature collected in real time, the irradiance and ambient temperature are divided into different types of working conditions, for each working condition, the physical mechanism model parameters are calculated and stored, the corresponding working condition is matched according to the irradiance and ambient temperature collected in real time, and the irradiance and ambient temperature are substituted into the physical mechanism model, and the theoretical value of photovoltaic output is output.
[0029] The deep learning model is specifically: Select the historical operation data, including: photovoltaic output value, irradiance, ambient temperature, relative humidity, wind speed, a total of 5-dimensional data, the time step is set to X minutes, 4C time step data is constructed as an input sequence, and the photovoltaic output value of the future C time steps is predicted; Setting the number of kernels to 64, the embedding layer output is 64x64, the attention weight of the feature dimension is calculated through 1x1 convolution, the spatial attention output is consistent with the embedding layer output layer, and the time dependence relationship is extracted through a single-layer LSTM, the attention weight of the time step dimension is calculated, and the attention weight of the feature dimension and the attention weight of the time step dimension are fused by layer-by-layer multiplication; The optimal fusion model of the deep learning model is determined through the validation set, and the photovoltaic output value of the future C time steps is predicted according to the optimal fusion model.
[0030] The specific process of dynamically allocating the weights of the physical mechanism model and the deep learning model through the Bayesian probability model is: According to the historical operation data of the photovoltaic power station, the fluctuation amplitude of the irradiance in the past N time steps is calculated, and the fluctuation amplitude of the irradiance is specifically: the irradiance change rate, the fluctuation amplitude of the ambient temperature in the past N time steps is calculated, and the fluctuation amplitude of the ambient temperature is specifically: the ambient temperature change amount; When the irradiance change rate is less than the irradiance change rate threshold, and the ambient temperature change amount is less than the ambient temperature change amount threshold, it is determined that the working condition of the power station is a meteorological stable working condition; When the irradiance change rate is greater than the irradiance change rate threshold, and the ambient temperature change amount is greater than the ambient temperature change amount threshold, it is determined that the working condition of the power station is a meteorological unstable working condition; In the meteorological stable working condition, the initial weight of the deep learning model is set to , the initial weight of the physical mechanism model is ; The sum of the initial weight of the deep learning model and the initial weight of the physical mechanism model is 1; In the meteorological unstable working condition, the initial weight of the deep learning model is set to The initial weight of the physical mechanism model is The sum of the initial weight of the deep learning model and the initial weight of the physical mechanism model is 1. In the initial weight range under the meteorological stable condition and the meteorological unstable condition, the fusion error is calculated by Bayes, the prediction error of the deep learning model and the prediction error of the physical mechanism model in the historical operation data under the meteorological stable condition and the meteorological unstable condition are respectively counted, the weight of the physical mechanism model and the deep learning model is dynamically allocated according to the prediction error of the deep learning model, the prediction error of the physical mechanism model and the fusion error, and the best fusion weight combination is output.
[0031] The specific process of outputting the short-term photovoltaic output prediction curve is: Based on the best fusion weight combination, the physical mechanism model and the deep learning model are fused to output the short-term photovoltaic output prediction curve.
[0032] S3: constructing an optimization target model, and constraining the target optimization model, generating a Pareto optimal solution set by using a third generation non-dominated sorting genetic algorithm, and outputting a short-term distributed power output plan, a storage charging and discharging curve and a flexible load adjustment instruction.
[0033] The specific process of constructing the optimization target model is: The optimization target model includes: a light rejection rate model, a power grid loss model and a storage charging and discharging cost model. The light rejection rate model is: Wherein, is a theoretical value of photovoltaic output, is an actual grid-connected power, is a time step, is a time step index. The power grid loss model is: Wherein, is a total number of lines, is a line current, is a line resistance, is a line total number index. The storage charging and discharging cost model is: Wherein, is a storage charging power, is a storage discharging power, , is a time-of-use electricity price, 0.25 is a time step length, is a time step, is a time step index.
[0034] The specific process of generating a Pareto optimal solution set by using a third generation non-dominated sorting genetic algorithm is: Based on the target optimization model, we can obtain the total number of time steps for the curtailment rate model, grid loss model, and energy storage charging and discharging cost model in the target optimization model, and set the population size according to the total number of time steps. By using non-dominated sorting to filter out the frontier solution set where no solution is comprehensively superior to itself, the crowding degree of solutions within the frontier is calculated to preserve diversity. Evolution is simulated by simulating binary crossover and polynomial mutation to generate offspring schemes. Then, with the abandonment rate approaching 0, network loss approaching 0, and cost approaching 0 as reference points, schemes that are close to the reference point and have high congestion are selected to form the next generation. This process is repeated until the target value change rate of the optimal scheme for 5 consecutive generations is <1%. Finally, a Pareto optimal solution set covering the multi-objective trade-off relationship is output.
[0035] The specific process of outputting the short-term distributed power generation plan, energy storage charging and discharging curve, and flexible load adjustment command is as follows: Based on the generated Pareto optimal solution set, solutions that meet preset conditions are selected from the Pareto optimal solution set, and the output includes short-term distributed power output plan: the future grid-connected power of photovoltaic power plants, energy storage charging and discharging curve: the time series of energy storage charging and discharging power; and flexible load adjustment instructions: the power adjustment amount of industrial load at each time step.
[0036] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A smart power integrated operation and management system, characterized in that: include: Data acquisition module: Acquires historical operating data of the photovoltaic power station; Hybrid prediction module: Constructs a physical mechanism model equivalent to a single diode and a spatiotemporal attention mechanism LSTM deep learning model. Combined with historical operation data of photovoltaic power plants, the weights of the physical mechanism model and the deep learning model are dynamically allocated through a Bayesian probability model to output a short-term photovoltaic power output prediction curve. Multi-objective optimization module: Constructs an optimization objective model and imposes constraints on the objective optimization model. It uses a third-generation non-dominated sorting genetic algorithm to generate a Pareto optimal solution set and outputs short-term distributed power generation output plan, energy storage charging and discharging curves, and flexible load adjustment instructions.
2. The intelligent power integrated operation and management system according to claim 1, characterized in that: The historical operating data of the photovoltaic power station specifically includes: Photovoltaic power plant parameter data and photovoltaic power plant environmental data; Photovoltaic power plant parameter data includes: hourly power generation, module temperature, inverter efficiency, and photovoltaic output value. Photovoltaic power plant environmental data includes: irradiance, cloud cover, wind speed, relative humidity, and ambient temperature.
3. The intelligent power integrated operation and management system according to claim 1, characterized in that: The specific physical mechanism model is as follows: The physical mechanism model is as follows: ,in, For photocurrent, This refers to the output voltage of the photovoltaic power station. It is the reverse saturation current. This is the equivalent series resistance of the cells in a photovoltaic power station. This refers to the output current of the photovoltaic power station. The ideality factor reflects the degree of non-ideality of the PN junction. Boltzmann's constant, Absolute temperature For electron charge, This is the equivalent parallel resistance of the cells in a photovoltaic power station.
4. The intelligent power integrated operation and management system according to claim 3, characterized in that: The function of the physical mechanism model is as follows: Based on the real-time collected irradiance and ambient temperature, different types of operating conditions are classified according to irradiance and ambient temperature. For each operating condition, the parameters of the physical mechanism model are calculated and stored. The corresponding operating conditions are matched according to the real-time collected irradiance and ambient temperature, and the irradiance and ambient temperature are substituted into the physical mechanism model to output the theoretical value of photovoltaic power output.
5. The intelligent power integrated operation and management system according to claim 1, characterized in that: The deep learning model is specifically: Historical operating data is selected, including five dimensions: photovoltaic power output, irradiance, ambient temperature, relative humidity, and wind speed. The time step is set to X minutes, and 4C time step data are constructed as input sequences to predict the photovoltaic power output in the next C time steps. With the number of kernels set to 64, the output of the embedding layer is 64×64. Attention weights in the feature dimension are calculated through 1×1 convolutions, and spatial attention is output in the same way as the output of the embedding layer. A single-layer LSTM with 64 units per layer is used to extract temporal dependencies and calculate attention weights in the time step dimension. The attention weights in the feature dimension and the attention weights in the time step dimension are fused by multiplying layer by layer. The optimal fusion model of the deep learning model is determined by the validation set, and the photovoltaic power output value is predicted for the next C time steps based on the optimal fusion model.
6. The intelligent power integrated operation and management system according to claim 1, characterized in that: The specific process of dynamically allocating the weights of the physical mechanism model and the deep learning model using the Bayesian probability model is as follows: Based on the historical operating data of the photovoltaic power station, the fluctuation range of irradiance over the past N time steps is calculated. The specific fluctuation range of irradiance is: irradiance change rate. The fluctuation range of ambient temperature over the past N time steps is calculated. The specific fluctuation range of ambient temperature is: ambient temperature change amount. When the rate of change of irradiance is less than the threshold for the rate of change of irradiance and the change in ambient temperature is less than the threshold for the change in ambient temperature, the power plant is determined to be in a meteorologically stable operating condition. When the rate of change of irradiance is greater than the threshold for the rate of change of irradiance, and the change in ambient temperature is greater than the threshold for the change in ambient temperature, the power plant is determined to be in a meteorologically unstable operating condition. Under stable meteorological conditions, let the initial weights of the deep learning model be... The initial weights of the physical mechanism model are ; The sum of the initial weights of the deep learning model and the initial weights of the physical mechanism model is 1. Under unstable meteorological conditions, let the initial weights of the deep learning model be... The initial weights of the physical mechanism model are ; The sum of the initial weights of the deep learning model and the initial weights of the physical mechanism model is 1. Within the initial weight range under both stable and unstable meteorological conditions, the fusion error is calculated using Bayesian methods. The prediction errors of the deep learning model and the physical mechanism model are statistically analyzed in historical operating data under both stable and unstable meteorological conditions. Based on the prediction errors of the deep learning model, the physical mechanism model, and the fusion error, the weights of the physical mechanism model and the deep learning model are dynamically allocated, and the optimal fusion weight combination is output.
7. The intelligent power integrated operation and management system according to claim 1, characterized in that: The specific process of outputting the short-term photovoltaic power output prediction curve is as follows: Based on the optimal combination of fusion weights, the physical mechanism model and the deep learning model are integrated to output a short-term photovoltaic power output prediction curve.
8. The intelligent power integrated operation and management system according to claim 1, characterized in that: The specific process of constructing the optimization target model is as follows: The optimization target models include: curtailment rate model, grid loss model, and energy storage charging and discharging cost model; The light rejection rate model is as follows: ,in, This is the theoretical value of photovoltaic power output. This represents the actual grid-connected power. For time steps, For time step index; The power grid loss model is as follows: ,in, The total number of lines, For line current, For line resistance, Index for the total number of lines; Energy storage charging and discharging cost model: ,in, For energy storage charging power, For energy storage discharge power, , This is for time-of-use pricing, with 0.25 as the time step. For time steps, For time step indexing.
9. The intelligent power integrated operation and management system according to claim 1, characterized in that: The specific process of generating the Pareto optimal solution set using the third-generation non-dominated sorting genetic algorithm is as follows: Based on the target optimization model, we can obtain the total number of time steps for the curtailment rate model, grid loss model, and energy storage charging and discharging cost model in the target optimization model, and set the population size according to the total number of time steps. By using non-dominated sorting to filter out the frontier solution set where no solution is comprehensively superior to itself, the crowding degree of solutions within the frontier is calculated to preserve diversity. Evolution is simulated by simulating binary crossover and polynomial mutation to generate offspring schemes. Then, with the abandonment rate approaching 0, network loss approaching 0, and cost approaching 0 as reference points, schemes that are close to the reference point and have high congestion are selected to form the next generation. This process is repeated until the target value change rate of the optimal scheme for 5 consecutive generations is <1%. Finally, a Pareto optimal solution set covering the multi-objective trade-off relationship is output.
10. The intelligent power integrated operation and management system according to claim 1, characterized in that: The specific process of outputting the short-term distributed power generation plan, energy storage charging and discharging curve, and flexible load adjustment command is as follows: Based on the generated Pareto optimal solution set, solutions that meet preset conditions are selected from the Pareto optimal solution set, and output includes short-term distributed power generation output plan, energy storage charging and discharging curve, and flexible load adjustment instructions.