A Time Series Simulation Method and System for New Energy Uncertainty Power Systems Based on GAN-RNN

Abstract

A time-series simulation method and system for uncertain new energy power systems based on GAN-RNN includes: acquiring historical meteorological data and power output data as the basic dataset; dividing temperature intervals and calculating the probability of occurrence of each temperature interval; calculating the scenario complexity; selecting typical monthly, typical weekly, and typical daily samples based on the probability of occurrence of temperature intervals and scenario complexity; constructing a power output data simulation model based on bidirectional RNN and GAN, and setting a physical rule layer; introducing a physical constraint regularization term into the generator loss function; training the model to obtain a trained model; inputting meteorological data of samples from each typical block into the trained model to generate power output sequences; calculating the corresponding system operation indicators based on the power output sequences corresponding to each typical block, and weighting and summing the system operation indicators based on the probability of occurrence of the temperature interval to which each typical block belongs to obtain the final operation indicators. This invention offers fast simulation speed and high accuracy.

Description

Technical Field

[0001] This invention belongs to the field of power system time-series operation simulation, and in particular relates to a time-series simulation method and system for new energy uncertain power systems based on GAN-RNN. Background Technology

[0002] With the accelerated transformation of the energy structure, the penetration rate of new energy sources such as wind power and photovoltaics in the power system continues to increase. Their output is highly volatile, random, and intermittent due to the influence of natural environmental factors (wind speed, irradiance, temperature, etc.), posing a severe challenge to the power system's "safe and stable operation, optimized economic dispatch, and balanced supply and demand regulation." Traditional approaches mainly rely on statistical methods such as probabilistic models, Monte Carlo sampling, or Markov chains to generate scenarios. However, these methods often struggle to capture high-dimensional time-series correlations and perform poorly under insufficient sample sizes or extreme weather conditions. Furthermore, traditional models make relatively coarse assumptions about the probability distribution of new energy output, limiting the reliability of simulation results.

[0003] Power system time-series simulation is a fundamental tool for the analysis and planning of power systems with a high proportion of renewable energy. Using a time step of 8760 hours (annual) or finer granularity, time-series simulation can simultaneously consider various constraints such as load, wind and solar power output, unit start-up and shutdown, energy storage dispatch, and transmission flow. It can more realistically reflect the impact of uncertainties in renewable energy output on the power system. However, existing power system time-series simulation tools still have two shortcomings: they cannot analyze planned wind and solar power projects, and their solution speed is slow when used in power systems with limited flexibility. These two shortcomings greatly limit the application scope of time-series simulation. Therefore, there is an urgent need to develop a power system time-series simulation method that, while considering the uncertainties in renewable energy output, ensures the accuracy of simulation results and accelerates the solution speed. Summary of the Invention

[0004] To address the limitations of traditional power system time-series operation simulation methods in analyzing planned wind and solar power projects and their slow solution speed when used in power systems with limited flexibility, this invention proposes a GAN-RNN-based power system time-series simulation method for new energy uncertainties. The aim is to provide the power grid with a power system time-series operation simulation method that accelerates the simulation process and achieves high accuracy while considering the uncertainty of new energy output.

[0005] The present invention adopts the following technical solution.

[0006] In a first aspect, this invention discloses a time-series simulation method for uncertain power systems with new energy sources based on GAN-RNN, the method comprising the following steps: Acquire historical meteorological data for the whole year, as well as corresponding new energy power output data and thermal power unit power output data at the same time, as the basic dataset; Temperature intervals are divided based on historical meteorological data throughout the year. The proportion of days in each temperature interval to the total number of days in the year is calculated as the probability of occurrence of that temperature interval. The scenario complexity is calculated based on the fluctuation range of new energy output within a single day. Typical month, typical week, and typical day samples are selected by combining the probability of occurrence of temperature intervals and the scenario complexity. A power output data simulation model based on bidirectional RNN and GAN is constructed. The model has a built-in physical rule layer for embedding power system engineering constraints. A physical constraint regularization term is introduced into the generator loss function of GAN. The physical constraint regularization term includes output upper and lower limit constraint loss, output change rate constraint loss, and new energy output physical mapping relationship constraint loss. The basic dataset is used as training samples to train the power output data simulation model, resulting in a trained model. Meteorological data from typical months, weeks, and days are input into the trained model to generate new energy output sequences and thermal power unit output sequences at the corresponding time scales. Based on the new energy power output sequence and thermal power unit power output sequence corresponding to typical months, typical weeks and typical days, the corresponding system operation indicators are calculated respectively. Then, the system operation indicators are weighted and summed based on the probability of occurrence of the temperature range to which the typical month, typical week and typical day belong, to obtain the final operation indicators.

[0007] More preferably, The temperature range division based on historical meteorological data throughout the year is specifically as follows: Based on annual temperature data, a sliding interval is divided according to a preset temperature interval width and a preset temperature interval overlap, so that adjacent temperature intervals have a preset overlap range, which is determined by the preset temperature interval overlap range.

[0008] More preferably, The complexity of the scenario is determined in the following manner:

[0009] in, For the first k The complexity of the scene on the day, For the first in the sample k The time step of the day, For the first k day t The rate of change of output at any given time; For all of the day The mean.

[0010] More preferably, The selection of typical monthly, weekly, and daily samples, based on the probability of occurrence of temperature ranges and the complexity of the scenario, includes the following steps: First, define high-complexity days: calculate the scene complexity of all sample days within the current temperature range to obtain a scene complexity set, sort the scene complexity set, and take the median as the scene complexity threshold; define sample days whose scene complexity is greater than or equal to the scene complexity threshold as high-complexity days. Secondly, considering the probability of a temperature range occurring and the sample selection for the high-complexity day: if the probability of a certain temperature range occurring... satisfy Then, 30 days are selected from the sample days corresponding to this interval to form a typical monthly sample, and the number of high-complexity days must be greater than or equal to 30 days. If the probability of occurrence of a certain temperature range satisfy Then, select 7 days from the sample days corresponding to this interval to form a typical weekly sample, and the number of high-complexity days must be greater than or equal to 7. If the probability of occurrence of a certain temperature range satisfy Then, one day is selected from the sample days corresponding to this interval as a typical day sample, preferably a high-complexity day; where, , For the preset probability threshold, , This is a preset quantity threshold.

[0011] More preferably, The construction of the power output data simulation model based on bidirectional RNN and GAN specifically includes: A hybrid GAN architecture integrating deep convolutional networks and bidirectional RNNs is constructed, the architecture including a generator, a discriminator, and a physical rule layer; In the generator, a bidirectional RNN is embedded after the deep convolutional layer and before the output layer as a temporal feature enhancement unit to enhance the temporal dependencies of the input features. In the discriminator, a bidirectional RNN is embedded after the deep convolutional layer and before the classification layer as a temporal consistency verification unit, used to verify the temporal rationality of the input sequence. The physical rules layer incorporates power system engineering constraints to ensure that the generated output sequence conforms to engineering realities.

[0012] More preferably, The generator loss function of the GAN is as follows:

[0013] in, The original GAN ​​adversarial loss; To constrain losses by upper and lower limits of output, Constraints on the rate of change of output, Constraints on the physical mapping relationship for new energy output; , , This is the weight for the regularization term, which is adjusted based on the actual training results.

[0014] More preferably, The output upper and lower limit constraint losses are specifically as follows:

[0015] in, , These are the lower and upper threshold values ​​for output, respectively; For the generated output sequence, the first... t Output value at any given moment The length of the output sequence; The output change rate constraint loss is specifically as follows:

[0016] in, For the maximum allowable rate of change of output, , The generated output sequence is respectively the first... t , t The output value at time -1.

[0017] More preferably, The aforementioned constraint loss on the physical mapping relationship of new energy output includes the constraint loss on the physical mapping relationship of wind power output. and the constraint loss of the physical mapping relationship of photovoltaic output Specifically:

[0018]

[0019] in, for t The wind power output value generated in real time; The theoretical output calculated based on wind speed and the turbine power curve; for t The photovoltaic output value generated at any time; The temperature coefficient of photovoltaic power. for t Ambient temperature at all times; Temperature under standard test conditions; This is the theoretical maximum output of the photovoltaic unit calculated based on the current light intensity.

[0020] More preferably, When training the power output data simulation model, a phased training method is adopted, which includes a pre-training phase and an adversarial training phase. During the pre-training phase, the weighted sum of the output upper and lower limit constraint loss and the output change rate constraint loss is used as the loss term to train the bidirectional RNN separately. During the adversarial training phase, the discriminator parameters in the GAN are updated based on cross-entropy loss; the generator parameters are updated based on the generator loss function.

[0021] Secondly, this invention discloses a time series simulation system for uncertain new energy power systems based on GAN-RNN, which is based on the aforementioned method. The system includes a basic data acquisition module, a typical sample screening module, a power output data simulation model construction module, a power output data simulation model training module, a power output sequence generation module, and a system operation index calculation module. The basic data acquisition module acquires historical meteorological data for the whole year, as well as the corresponding new energy power output data and thermal power unit power output data at the same time, as the basic dataset; The typical sample selection module divides temperature intervals based on historical meteorological data throughout the year, calculates the proportion of days in each temperature interval to the total number of days in the year as the probability of occurrence of that temperature interval, calculates the scenario complexity based on the fluctuation range of new energy output within a single day, and selects typical monthly, typical weekly, and typical daily samples by combining the probability of occurrence of temperature intervals and scenario complexity. The power output data simulation model construction module constructs a power output data simulation model based on bidirectional RNN and GAN. The model has a built-in physical rule layer for embedding power system engineering constraints; and introduces physical constraint regularization terms into the generator loss function of GAN. The physical constraint regularization terms include output upper and lower limit constraint loss, output change rate constraint loss, and new energy output physical mapping relationship constraint loss. The power output data simulation model training module uses the basic dataset as training samples to train the power output data simulation model and obtain the trained model. The power output sequence generation module inputs meteorological data of typical monthly, typical weekly, and typical daily samples into the trained model to generate new energy power output sequences and thermal power unit power output sequences at the corresponding time scales. The system operation index calculation module calculates the corresponding system operation indexes based on the new energy output sequence and thermal power unit output sequence corresponding to typical months, typical weeks, and typical days. It also weights and superimposes the system operation indexes based on the probability of occurrence of the temperature range to which the typical month, typical week, and typical day belong, to obtain the final operation indexes.

[0022] The beneficial effects of this invention are compared with those of the prior art: Existing technologies often directly use raw data or random sampling, and their scenario selection relies heavily on experience or single probability indicators, resulting in insufficient sample representativeness or excessive computational load, making it difficult to balance efficiency and accuracy. This invention proposes a hierarchical selection strategy based on temperature range probability and scenario complexity. High-probability ranges are covered by typical months, medium-probability ranges by typical weeks, and low-probability ranges by typical days, focusing on extreme scenarios. This approach ensures coverage of major operating scenarios throughout the year while highlighting the impact of high-complexity, low-probability events, significantly reducing the sample size for subsequent indicator calculations and significantly improving analysis efficiency. Traditional time-series simulation models struggle to accurately depict the strong volatility, nonlinearity, and temporal dependencies of renewable energy output, especially exhibiting significant errors in extreme scenarios. This invention employs a hybrid architecture combining bidirectional RNNs and GANs, which can accurately model the temporal dependencies of output and generate volatile sequences that conform to the real distribution. Combined with the feature extraction capabilities of deep convolution, it can more accurately capture the dynamic changes in renewable energy output, particularly the volatility characteristics under high-complexity scenarios. Furthermore, in existing technologies, the output sequences generated by traditional GAN ​​models often deviate from the physical constraints of the power system, making the simulation results unsuitable for direct engineering analysis. This invention introduces multi-dimensional physical constraint regularization terms into the generator loss function and employs hard correction at the physical rule layer to ensure that the generated output sequences strictly conform to engineering realities, significantly improving the reliability and practicality of the simulation results.

[0023] Existing technologies often rely on massive amounts of raw data for calculation, which is computationally intensive, time-consuming, and makes it difficult to focus on key scenarios. This invention replaces full-time simulation with typical block overlay. It only needs to be based on the simulation results of typical months, weeks, and days to derive comprehensive indicators that reflect the characteristics of the whole year, which greatly improves the evaluation efficiency and is especially suitable for rapid analysis in large-scale new energy grid connection scenarios. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the time-series simulation method for new energy uncertain power systems based on GAN-RNN according to the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0026] like Figure 1As shown, this invention discloses a time-series simulation method for uncertain power systems in new energy sources based on GAN-RNN. The method includes the following steps: Step 1: Obtain historical meteorological data for the whole year, as well as the corresponding new energy power output data and thermal power unit power output data at the same time, as the basic dataset; Step 2: Divide temperature intervals based on historical meteorological data for the whole year, and calculate the proportion of days in each temperature interval to the total number of days in the year as the probability of occurrence of that temperature interval; calculate the scenario complexity based on the fluctuation range of new energy output within a single day; select typical month, typical week and typical day samples by combining the probability of occurrence of temperature intervals and scenario complexity. The temperature range division based on historical meteorological data throughout the year is specifically as follows: Based on annual temperature data, a sliding interval is divided according to a preset temperature interval width and a preset temperature interval overlap, so that adjacent temperature intervals have a preset overlap range, which is determined by the preset temperature interval overlap range.

[0027] The complexity of the scenario is determined in the following manner:

[0028] in, For the first k The complexity of the scene on the day, For the first in the sample k The time step of the day, For the first k day t The rate of change of output at any given time; For all of the day The mean.

[0029] The selection of typical monthly, weekly, and daily samples, based on the probability of occurrence of temperature ranges and the complexity of the scenario, includes the following steps: First, define high-complexity days: calculate the scene complexity of all sample days within the current temperature range to obtain a scene complexity set, sort the scene complexity set, and take the median as the scene complexity threshold; define sample days whose scene complexity is greater than or equal to the scene complexity threshold as high-complexity days. Secondly, considering the probability of a temperature range occurring and the sample selection for the high-complexity day: if the probability of a certain temperature range occurring... satisfy Then, 30 days are selected from the sample days corresponding to this interval to form a typical monthly sample, and the number of high-complexity days must be greater than or equal to 30 days. If the probability of occurrence of a certain temperature range satisfy Then, select 7 days from the sample days corresponding to this interval to form a typical weekly sample, and the number of high-complexity days must be greater than or equal to 7. If the probability of occurrence of a certain temperature range satisfy Then, one day is selected from the sample days corresponding to this interval as a typical day sample, preferably a high-complexity day; where, , For the preset probability threshold, , This is a preset quantity threshold.

[0030] Step 3: Construct a power output data simulation model based on bidirectional RNN and GAN. The model has a built-in physical rule layer to embed power system engineering constraints; and introduces a physical constraint regularization term into the generator loss function of GAN. The physical constraint regularization term includes output upper and lower limit constraint loss, output change rate constraint loss, and new energy output physical mapping relationship constraint loss. The construction of the power output data simulation model based on bidirectional RNN and GAN specifically includes: A hybrid GAN architecture integrating deep convolutional networks and bidirectional RNNs is constructed, the architecture including a generator, a discriminator, and a physical rule layer; In the generator, a bidirectional RNN is embedded after the deep convolutional layer and before the output layer as a temporal feature enhancement unit to enhance the temporal dependencies of the input features. In the discriminator, a bidirectional RNN is embedded after the deep convolutional layer and before the classification layer as a temporal consistency verification unit, used to verify the temporal rationality of the input sequence. The physical rules layer incorporates power system engineering constraints to ensure that the generated output sequence conforms to engineering realities.

[0031] The generator loss function of the GAN is as follows:

[0032] in, The original GAN ​​adversarial loss; To constrain losses by upper and lower limits of output, Constraints on the rate of change of output, Constraints on the physical mapping relationship for new energy output; , , This is the weight for the regularization term, which is adjusted based on the actual training results.

[0033] The output upper and lower limit constraint losses are specifically as follows:

[0034] in, , These are the lower and upper threshold values ​​for output, respectively; For the generated output sequence, the first... t Output value at any given moment The length of the output sequence; The output change rate constraint loss is specifically as follows:

[0035] in, For the maximum allowable rate of change of output, , The generated output sequence is respectively the first... t , t The output value at time -1.

[0036] The aforementioned constraint loss on the physical mapping relationship of new energy output includes the constraint loss on the physical mapping relationship of wind power output. and the constraint loss of the physical mapping relationship of photovoltaic output Specifically:

[0037]

[0038] in, for t The wind power output value generated in real time; The theoretical output calculated based on wind speed and the turbine power curve; for t The photovoltaic output value generated at any time; The temperature coefficient of photovoltaic power. for t Ambient temperature at all times; Temperature under standard test conditions; This is the theoretical maximum output of the photovoltaic unit calculated based on the current light intensity.

[0039] Step 4: Use the basic dataset as training samples to train the power output data simulation model to obtain the trained model; When training the power output data simulation model, a phased training method is adopted, which includes a pre-training phase and an adversarial training phase. During the pre-training phase, the weighted sum of the output upper and lower limit constraint loss and the output change rate constraint loss is used as the loss term to train the bidirectional RNN separately. During the adversarial training phase, the discriminator parameters in the GAN are updated based on cross-entropy loss; the generator parameters are updated based on the generator loss function.

[0040] Step 5: Input the meteorological data of typical monthly, typical weekly, and typical daily samples into the trained model to generate the new energy output sequence and thermal power unit output sequence at the corresponding time scale; Step 6: Based on the new energy output sequence and thermal power unit output sequence corresponding to typical month, typical week and typical day, calculate the corresponding system operation indicators respectively, and perform weighted summation of each system operation indicator based on the occurrence probability of the temperature range to which the typical month, typical week and typical day belong, to obtain the final operation indicators.

[0041] The present invention also discloses a time series simulation system for new energy uncertain power systems based on GAN-RNN based on the aforementioned method, including a basic data acquisition module, a typical sample screening module, a power output data simulation model construction module, a power output data simulation model training module, a power output sequence generation module, and a system operation index calculation module; The basic data acquisition module acquires historical meteorological data for the whole year, as well as the corresponding new energy power output data and thermal power unit power output data at the same time, as the basic dataset; The typical sample selection module divides temperature intervals based on historical meteorological data throughout the year, calculates the proportion of days in each temperature interval to the total number of days in the year as the probability of occurrence of that temperature interval, calculates the scenario complexity based on the fluctuation range of new energy output within a single day, and selects typical monthly, typical weekly, and typical daily samples by combining the probability of occurrence of temperature intervals and scenario complexity. The power output data simulation model construction module constructs a power output data simulation model based on bidirectional RNN and GAN. The model has a built-in physical rule layer for embedding power system engineering constraints; and introduces physical constraint regularization terms into the generator loss function of GAN. The physical constraint regularization terms include output upper and lower limit constraint loss, output change rate constraint loss, and new energy output physical mapping relationship constraint loss. The power output data simulation model training module uses the basic dataset as training samples to train the power output data simulation model and obtain the trained model. The power output sequence generation module inputs meteorological data of typical monthly, typical weekly, and typical daily samples into the trained model to generate new energy power output sequences and thermal power unit power output sequences at the corresponding time scales. The system operation index calculation module calculates the corresponding system operation indexes based on the new energy output sequence and thermal power unit output sequence corresponding to typical months, typical weeks, and typical days. It also weights and superimposes the system operation indexes based on the probability of occurrence of the temperature range to which the typical month, typical week, and typical day belong, to obtain the final operation indexes.

[0042] Example 1: like Figure 1As shown, this invention discloses a time-series simulation method for uncertain power systems in new energy sources based on GAN-RNN. The method includes the following steps: Step 1: Obtain historical meteorological data for the whole year, as well as the corresponding new energy power output data and thermal power unit power output data at the same time, as the basic dataset; Step 2: Divide temperature intervals based on historical meteorological data for the whole year, and calculate the proportion of days in each temperature interval to the total number of days in the year as the probability of occurrence of that temperature interval; calculate the scenario complexity based on the fluctuation range of new energy output within a single day; select typical month, typical week and typical day samples by combining the probability of occurrence of temperature intervals and scenario complexity. Step 2.1: Based on the annual temperature data, perform sliding interval division according to the preset temperature interval width and preset temperature interval overlap, so that adjacent temperature intervals have a preset overlap range, which is determined by the preset temperature interval overlap range. Specifically, in this embodiment, the preset temperature range width is 6°C and the preset temperature range overlap is 1°C; the annual temperature data provided by NASA is divided according to the rule of "6°C interval, 1°C overlap", such as [0, 6], [6, 12]....

[0043] Take the average temperature over a single day (24 hours). If the average temperature of a day falls within a certain temperature range (e.g., 10-16℃), then that day is included in the corresponding temperature range.

[0044] The actual number of days corresponding to each temperature range throughout the year is counted, and the probability of occurrence of each temperature range is calculated using the formula "number of days in the range / total number of days in the year".

[0045] Step 2.2: The computational scenario is complex; The complexity of the scenario is determined in the following manner:

[0046] in, For the first k The complexity of the scene on the day, For the first in the sample k The time step of the day, For the first k day t The rate of change of output at any given time; For all of the day The mean.

[0047] Step 2.3: Select typical month, typical week, and typical day samples based on the probability of occurrence of temperature ranges and the complexity of the scene; First, define a high-complexity day: calculate the scenario complexity of all sample days within the current temperature range to obtain the scenario complexity set. Where K is the total number of days in the year for that interval; sort the set of scene complexity and take the median as the scene complexity threshold. The scene complexity must be greater than or equal to the scene complexity threshold. The sample day is defined as a high-complexity day; Secondly, considering the probability of a temperature range occurring and the sample selection for the high-complexity day: if the probability of a certain temperature range occurring... satisfy Then, 30 days are selected from the sample days corresponding to this interval to form a typical monthly sample, and the number of high-complexity days must be greater than or equal to 30 days. If the probability of occurrence of a certain temperature range satisfy Then, select 7 days from the sample days corresponding to this interval to form a typical weekly sample, and the number of high-complexity days must be greater than or equal to 7. If the probability of occurrence of a certain temperature range satisfy Then, one day is selected from the sample days corresponding to this interval as a typical day sample, preferably a high-complexity day; where, , For the preset probability threshold, , This is a preset quantity threshold.

[0048] Preferably, , It can be determined in the following way: First, calculate the probability of all temperature ranges and sort them from largest to smallest. Then, determine the threshold by referring to the percentage quantile rule commonly used in climate statistics. For example, set the 75th percentile of the sorted probabilities as... The 25% quantile is set as Alternatively, we can draw on the approach to defining extreme weather events. If it is necessary to emphasize the impact of high-frequency temperature ranges, the 80th percentile can be set as... The 30% quantile is set as .

[0049] It is usually between 0.15 and 0.3 (corresponding to about 55-110 days in a year, which is consistent with the data source size of a typical 30-day month). It is usually between 0.03 and 0.08 (corresponding to about 11-30 days a year, which can support the selection of typical 7-day week samples and also distinguish low probability intervals).

[0050] , The coverage ratio can be determined comprehensively based on actual power grid operation statistics and the minimum coverage ratio in high-complexity scenarios. In this embodiment, Preferably 9, The preferred value is 2.

[0051] Step 3: Construct a power output data simulation model based on bidirectional RNN and GAN. The model has a built-in physical rule layer to embed power system engineering constraints; and introduces a physical constraint regularization term into the generator loss function of GAN. The physical constraint regularization term includes output upper and lower limit constraint loss, output change rate constraint loss, and new energy output physical mapping relationship constraint loss. Specifically, it includes: Step 3.1: Construct a hybrid GAN architecture integrating deep convolutional networks and bidirectional RNNs. The architecture includes a generator, a discriminator, and a physical rule layer. In the generator, the bidirectional RNN is embedded after the deep convolutional layer and before the output layer as a temporal feature enhancement unit to strengthen the temporal dependencies of the input features. In the discriminator, the bidirectional RNN is embedded after the deep convolutional layer and before the classification layer as a temporal consistency verification unit to verify the temporal rationality of the input sequence. The physical rule layer incorporates power system engineering constraints to ensure that the generated output sequence conforms to engineering realities.

[0052] Specifically, this invention constructs a hybrid GAN architecture that integrates deep convolutional networks and bidirectional RNNs. Its core consists of two main modules: a generator (G) and a discriminator (D). The RNN is embedded in the GAN framework in a dual role of "feature extraction and temporal constraints." The overall architecture is as follows: Generator: Deep convolutional layer (feature mapping) + bidirectional RNN layer (temporal dependency modeling) + output layer (generation of new energy power output sequence); Discriminator: Deep convolutional layer (feature extraction) + bidirectional RNN layer (temporal consistency verification) + output layer (real / fake sequence discrimination); Physical rules layer: Built-in power system engineering constraints (output constraints and the mapping relationship between wind speed / light intensity and output) to ensure that the generated sequence conforms to the actual engineering situation.

[0053] The embedding positions and input / output of a bidirectional RNN are as follows: 1) Bidirectional RNN in the generator Embedding location: after the deep convolutional layer and before the output layer, as a temporal feature enhancement unit; Input: High-dimensional feature vectors extracted by the convolutional layer, containing fused features of historical new energy output and meteorological forecast data; Output: An intermediate feature vector with temporal dependencies (with the same dimension as the input), which is passed to the output layer to generate a new energy output sequence with continuous time steps.

[0054] 2) Bidirectional RNN in the discriminator Embedding location: after the deep convolutional layer and before the classification layer, serving as a temporal consistency verification unit; Input: Real new energy output sequence or pseudo sequence output by generator (after feature extraction by convolutional layer); Output: A dynamic score vector of time-series features, used to assist the discriminator in judging the authenticity of the input sequence (final output is a single probability value, between 0 and 1).

[0055] Specifically, the deep convolutional layer uses 3 convolutional layers and pooling operations, with a kernel size of 3×3 and the activation function being LeakyReLU, to extract spatial features and local dependencies of the data; The bidirectional RNN layer has 256 hidden neurons and uses the tanh activation function to capture both forward and backward dependencies of the sequence, thus solving the long-term dependency forgetting problem of traditional RNNs. The physical rules layer is based on the principles of new energy generators and includes built-in output upper limit constraints (wind turbine rated power, photovoltaic module maximum output) and rate of change constraints (output surge and drop thresholds). The constraints in the physical rule layer specifically include: (1) Wind power output constraints Output upper and lower limit constraints:

[0056] in, This refers to the rated power of the fan. The minimum technical output of the wind turbine (usually 5% to 10% of the rated power, i.e.) = 0.05 P w，max ~ 0.1 P w，max ); P w，t for t The wind power output value generated at any time.

[0057] Output change rate constraint:

[0058] in, This represents the maximum allowable rate of change in wind turbine output (usually 10%~15% of rated power / 15min), where 15min is the data time granularity. If other granularities are used, adjustments need to be made proportionally (e.g., for a 5min granularity, use...). / 3).

[0059] Wind speed-output mapping constraint: Piecewise function constraint based on wind turbine power curve. Let the wind speed be... ,but: when or hour, ; when hour, , is a polynomial fitting of the power curve of the rising section of the fan, and the coefficients a, b, c, d are determined by the manufacturer's parameters for the specific fan model; when hour, ; in, The cut-in wind speed is typically 3-4 m / s; To cut off the wind speed, it is usually 25~30m / s; Rated wind speed; (2) Photovoltaic output constraints Output upper and lower limit constraints:

[0060] in, for t Solar power output at all times; Rated power of photovoltaic modules; for t Solar irradiance at any time (W / m 2 ), Let be the light intensity correction factor, and satisfy . , Irradiance under standard test conditions; mandatory constraints during nighttime hours (solar altitude angle < 0). .

[0061] Output change rate constraint:

[0062] in, This is the maximum allowable rate of change in photovoltaic output (usually taken as 20%~30% of rated power / 15min).

[0063] Temperature correction constraints:

[0064] in, Photovoltaic output without considering the effect of temperature; The photovoltaic power temperature coefficient (typically -0.002 to -0.005 / ℃); for t Ambient temperature at all times; The temperature is the standard test temperature (usually 25°C).

[0065] (3) Output constraints of thermal power units Output upper and lower limit constraints:

[0066] in, Let t be the output of the thermal power unit; Rated power of thermal power unit; Minimum technical output (30% to 50% of rated power for coal-fired units and 10% to 20% for gas-fired units).

[0067] Climbing speed constraint:

[0068] in, The maximum load increase rate (typically 1% to 2% for coal-fired units) Gas turbine units typically account for 3% to 5% of this. ), This is the maximum rate of load reduction (usually equal to or slightly greater than the rate of load increase).

[0069] Start-stop constraints: If t If the unit is in the startup state (changing from shutdown to operation), then It needs to be gradually increased from 0 to And startup time satisfy , For the shortest start-up time, coal-fired units typically require 2-4 hours, while gas-fired units require 10-30 minutes; if t If the unit is in a shutdown state (changing from operation to shutdown), then The output needs to be gradually reduced to 0, and the downtime should be reduced. satisfy , This is the shortest downtime, usually 1 to 2 hours.

[0070] Specifically, the physical rule layer embedding implementation method is as follows: A constraint verification and correction sub-network is added after the generator output layer to correct the generated output sequence in real time and ensure that hard constraints are met. The modified subnetwork structure includes two fully connected layers (128 hidden neurons, ReLU activation function) and an output layer (Sigmoid activation function for normalization correction). The specific correction logic is as follows: The original output of the input generator at the physical rule layer Influencing factors (wind speed) at the corresponding time Irradiance ,temperature ); The allowable output range is calculated based on the constraint rules in the aforementioned physical rule layer. With allowable variation ,like Then it is corrected to ;like Then it is corrected to ;like Then it is corrected to ;like Then it is corrected to ;in, , These are the upper and lower limits of the allowable output, respectively; , These are the upper limits of the rate of change of output; It is a local minimum of 1e-6, used to avoid gradient vanishing.

[0071] For constraints related to influencing factors such as wind speed-output and temperature correction, the corrected output is directly calculated through the correction sub-network. This ensures that the mapping relationship conforms to physical laws.

[0072] Step 3.2: Introduce a physical constraint regularization term into the generator loss function of the GAN. Gradient descent forces the model to learn an output sequence that conforms to the constraints. The generator loss function is as follows:

[0073] in, The original GAN ​​adversarial loss; To constrain losses by upper and lower limits of output, Constraints on the rate of change of output, Constraints on the physical mapping relationship for new energy output; , , This is the weight for the regularization term, adjusted based on actual training results, with values ​​ranging from 0.1 to 1.0 (recommended initial value is...). ).

[0074] The original GAN ​​adversarial loss Specifically:

[0075] in, z It is a random noise vector; G ( z D(·) represents the generator output; D(·) represents the discriminator output probability. The output upper and lower limit constraint loss Specifically:

[0076] in, T The length of the output sequence. P t For the generated output sequence, the first... t Output value at any given moment; , These are the lower and upper thresholds for output, respectively.

[0077] The output change rate constraint loss Specifically:

[0078] in, For the maximum allowable rate of change of output, , The generated output sequence is respectively the first... t , t The output value at time -1.

[0079] The aforementioned constraint loss on the physical mapping relationship of new energy output includes the constraint loss on the physical mapping relationship of wind power output. and the constraint loss of the physical mapping relationship of photovoltaic output Specifically:

[0080]

[0081]

[0082] in, for t The wind power output value generated in real time; The theoretical output calculated based on wind speed and the turbine power curve; for t The photovoltaic output value generated at any time; The temperature coefficient of photovoltaic power. for t Ambient temperature at all times; Temperature under standard test conditions; This is the theoretical maximum output of the photovoltaic unit calculated based on the current light intensity; for t Solar irradiance at any given time; Irradiance under standard test conditions.

[0083] Thermal power units have no physical mapping relationship constraint loss; their start-up and shutdown constraints are achieved through modified subnetworks in the network structure.

[0084] Step 4: Use the basic dataset as training samples to train the power output data simulation model to obtain the trained model; Specifically, it includes: Step 4.1: Data preprocessing; To input the model, a standardized time-series dataset is constructed, collecting time-series data such as active / reactive power and power factor of wind turbines, photovoltaics, and thermal power (usually at a granularity of 5-15 minutes or 1 hour); and collecting the influencing factors corresponding to the power data at the same time, including meteorological data such as wind speed, irradiance, ambient temperature, and cloud cover. Outlier removal: Using the 3σ principle, box plot method, etc., remove extreme values ​​in the output data caused by equipment failure and measurement error (such as wind turbine output suddenly dropping to 0 with no shutdown record, photovoltaic output abnormally not being 0 at night, etc.). Imputing missing values: For data gaps (such as 1-2 hours of missing data due to sensor failure), linear interpolation, LSTM interpolation, or neighborhood mean imputation are used to fill in the gaps and avoid breaks in the continuity of time sequence. Standardization processing: Perform Min-Max standardization or Z-Score standardization on all input data (output + influence factor) to eliminate dimensional differences; Time series window construction: The standardized data is divided into time series sample windows of fixed length according to the time step. Each window must contain "influence factor sequence + corresponding output sequence" to form a 3D tensor of [number of samples, time step, feature dimension]. Data partitioning: Divide the time series samples into a training set (for model parameter learning), a validation set (for adjusting hyperparameters and avoiding overfitting), and a test set (for evaluating the model's generalization ability) in a ratio of 7:2:1 or 8:1:1.

[0085] Step 4.2: Adopt a phased training approach, specifically including a pre-training phase and an adversarial training phase; in the pre-training phase, use the sum of the output upper and lower limit constraint loss and the output change rate constraint loss as the loss term to train the bidirectional RNN separately; in the adversarial training phase, update the discriminator parameters in the GAN based on the cross-entropy loss; update the generator parameters based on the generator loss function. Specifically, during the pre-training phase, only [the following is enabled] and Regular terms, Set the weight to 0.8 (higher weight) to force the model to initially learn the constraints of the output range and rate of change; During the adversarial training phase, the following was introduced: Regular terms, Gradually increase from 0.1 to 0.4, while adjusting... To the recommended value, achieve collaborative optimization of adversarial generation and physical constraints; During training, the constraint satisfaction rate is monitored in real time. The constraint satisfaction rate is defined as: model output power. Falling within the allowed range Within, and the difference between the output power at the previous moment and the output power at the previous moment satisfies The percentage of training time that is used for training. Training can only be stopped when this percentage reaches 99% or higher, in order to ensure the engineering usability of the model output.

[0086] Specifically, In the pre-training stage: the bidirectional RNN in the generator is trained using real time-series data, with the goal of minimizing the prediction error (MSE loss) so that it can initially grasp the time-series variation pattern of new energy output; the bidirectional RNN in the discriminator is trained using real sequences and randomly generated pseudo sequences so that it can initially have the ability to distinguish between real and pseudo sequences; after iterating for a preset number of rounds, the parameters are fixed. In this embodiment, the preset number of rounds is preferably 100.

[0087] During the adversarial training phase: random noise and historical data features are fused and input into the generator to generate a pseudo-sequence of new energy output; the real sequence (label 1) and the pseudo-sequence (label 0) are input into the discriminator, and the temporal consistency is verified by a bidirectional RNN, and the discrimination result is output. The discriminator parameters are updated based on the cross-entropy loss (the learning rate is preferably 0.0002); the discriminator parameters are fixed, the generated pseudo-sequence is input into the discriminator, and the generator parameters are updated based on the generator loss function.

[0088] Iterative optimization: The generator and discriminator are trained alternately, and the learning rate is adjusted every 200 iterations (decay coefficient 0.95) until the loss function converges (generator loss is less than 0.1 and discriminator loss is close to 0.5).

[0089] Step 5: Input the meteorological data of typical monthly, typical weekly, and typical daily samples into the trained model to generate the new energy output sequence and thermal power unit output sequence at the corresponding time scale; Generator input construction: The meteorological data corresponding to typical monthly, typical weekly, and typical daily samples are standardized and used as an influencing factor sequence, which is then concatenated with a random noise vector and used as the input to the generator; Random noise vectors are sampled from a normal distribution with the same dimensions as during training (e.g., 100 dimensions) to ensure the diversity of the generated data; Simulated output generation and inverse standardization: The constructed input is fed into the trained generator, which outputs a standardized output sequence of [number of samples, time step, 1]. Inverse standardization is performed using the standardization parameters during training (such as the maximum and minimum values ​​of Min-Max) to map the output sequence from the [0,1] interval back to the actual physical units (MW) to obtain the final simulated output data.

[0090] Step 6: Based on the new energy output sequence and thermal power unit output sequence corresponding to typical month, typical week and typical day, calculate the corresponding system operation indicators respectively, and perform weighted summation of each system operation indicator based on the occurrence probability of the temperature range to which the typical month, typical week and typical day belong, to obtain the final operation indicators; The system operation indicators include: Power supply reliability indicators: Annual average outage time (SAIDI), annual average number of outages (SAIFI), and percentage of time spent operating without load gaps; Safety constraint satisfaction rate: percentage of overload duration of transmission lines throughout the year, percentage of overload duration of node voltage, and percentage of overload duration of frequency deviation. Flexibility reserve adequacy ratio: the percentage of time during which the fluctuations of new energy sources exceed the regulation capacity of thermal power throughout the year, and the number of times the reserve capacity is insufficient under extreme scenarios; Power utilization efficiency: annual utilization hours of wind power, annual utilization hours of photovoltaic power, and annual utilization hours of thermal power units; Related to wind and solar curtailment: Annual wind curtailment rate = Annual wind curtailment volume / Annual theoretical wind power generation; Annual solar curtailment rate = Annual solar curtailment volume / Annual theoretical photovoltaic power generation. Operating costs: total annual fuel cost, total start-up and shutdown cost of thermal power units, and average operating cost per unit of electricity generated across the entire grid.

[0091] Calculate the system operation indicators for the new energy power output sequence and thermal power unit power output sequence scenarios corresponding to typical months, typical weeks, and typical days, respectively; The specific calculation process for the final operating indicators is as follows: The system operating indicators for typical months, typical weeks, and typical days are calculated respectively; the probability of occurrence of the temperature range belonging to the typical month is multiplied by the system operating indicator calculated for the typical month to obtain the weighted operating indicator for the typical month; the probability of occurrence of the temperature range belonging to the typical week is multiplied by the system operating indicator calculated for the typical week to obtain the weighted operating indicator for the typical week; the probability of occurrence of the temperature range belonging to the typical day is multiplied by the system operating indicator calculated for the typical day to obtain the weighted operating indicator for the typical day; the weighted operating indicators for the typical month, the typical week, and the typical day are summed to obtain the final operating indicator reflecting the comprehensive operating characteristics throughout the year.

[0092] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0093] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0094] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0095] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A time-series simulation method for uncertain power systems in new energy sources based on GAN-RNN, characterized in that, The method includes the following steps: Acquire historical meteorological data for the whole year, as well as corresponding new energy power output data and thermal power unit power output data at the same time, as the basic dataset; Temperature intervals are divided based on historical meteorological data throughout the year. The proportion of days in each temperature interval to the total number of days in the year is calculated as the probability of occurrence of that temperature interval. The scenario complexity is calculated based on the fluctuation range of new energy output within a single day. Typical month, typical week, and typical day samples are selected by combining the probability of occurrence of temperature intervals and the scenario complexity. A power output data simulation model based on bidirectional RNN and GAN is constructed. The model has a built-in physical rule layer for embedding power system engineering constraints. A physical constraint regularization term is introduced into the generator loss function of GAN. The physical constraint regularization term includes output upper and lower limit constraint loss, output change rate constraint loss, and new energy output physical mapping relationship constraint loss. The basic dataset is used as training samples to train the power output data simulation model, resulting in a trained model. Meteorological data from typical months, weeks, and days are input into the trained model to generate new energy output sequences and thermal power unit output sequences at the corresponding time scales. Based on the new energy power output sequence and thermal power unit power output sequence corresponding to typical months, typical weeks and typical days, the corresponding system operation indicators are calculated respectively. Then, the system operation indicators are weighted and summed based on the probability of occurrence of the temperature range to which the typical month, typical week and typical day belong, to obtain the final operation indicators.

2. The time-series simulation method for uncertain new energy power systems according to claim 1, characterized in that: The temperature range division based on historical meteorological data throughout the year is specifically as follows: Based on annual temperature data, a sliding interval is divided according to a preset temperature interval width and a preset temperature interval overlap, so that adjacent temperature intervals have a preset overlap range, which is determined by the preset temperature interval overlap range.

3. The time-series simulation method for uncertain new energy power systems according to claim 1, characterized in that: The complexity of the scenario is determined in the following manner: in, For the first k The complexity of the scene on the day, For the first in the sample k The time step of the day, For the first k day t The rate of change of output at any given time; For all of the day The mean.

4. The time-series simulation method for new energy uncertain power systems according to claim 3, characterized in that: The selection of typical monthly, weekly, and daily samples, based on the probability of occurrence of temperature ranges and the complexity of the scenario, includes the following steps: First, define high-complexity days: calculate the scene complexity of all sample days within the current temperature range to obtain a scene complexity set, sort the scene complexity set, and take the median as the scene complexity threshold; define sample days whose scene complexity is greater than or equal to the scene complexity threshold as high-complexity days. Secondly, considering the probability of a temperature range occurring and the sample selection for the high-complexity day: if the probability of a certain temperature range occurring... satisfy Then, 30 days are selected from the sample days corresponding to this interval to form a typical monthly sample, and the number of high-complexity days must be greater than or equal to 30 days. If the probability of occurrence of a certain temperature range satisfy Then, select 7 days from the sample days corresponding to this interval to form a typical weekly sample, and the number of high-complexity days must be greater than or equal to 7. If the probability of occurrence of a certain temperature range satisfy Then, one day is selected from the sample days corresponding to this interval as a typical day sample, preferably a high-complexity day; where, , For the preset probability threshold, , This is a preset quantity threshold.

5. The time-series simulation method for uncertain new energy power systems according to claim 1, characterized in that: The construction of the power output data simulation model based on bidirectional RNN and GAN specifically includes: A hybrid GAN architecture integrating deep convolutional networks and bidirectional RNNs is constructed, the architecture including a generator, a discriminator, and a physical rule layer; In the generator, a bidirectional RNN is embedded after the deep convolutional layer and before the output layer as a temporal feature enhancement unit to enhance the temporal dependencies of the input features. In the discriminator, a bidirectional RNN is embedded after the deep convolutional layer and before the classification layer as a temporal consistency verification unit, used to verify the temporal rationality of the input sequence. The physical rules layer incorporates power system engineering constraints to ensure that the generated output sequence conforms to engineering realities.

6. The time-series simulation method for uncertain new energy power systems according to claim 5, characterized in that: The generator loss function of the GAN is as follows: in, The original GAN ​​adversarial loss; To constrain losses by upper and lower limits of output, Constraints on the rate of change of output, Constraints on the physical mapping relationship for new energy output; , , This is the weight for the regularization term, which is adjusted based on the actual training results.

7. The time-series simulation method for uncertain new energy power systems according to claim 6, characterized in that: The output upper and lower limit constraint losses are specifically as follows: in, , These are the lower and upper threshold values ​​for output, respectively; For the generated output sequence, the first... t Output value at any given moment The length of the output sequence; The output change rate constraint loss is specifically as follows: in, For the maximum allowable rate of change of output, , The generated output sequence is respectively the first... t , t The output value at time -1.

8. The time-series simulation method for uncertain new energy power systems according to claim 6, characterized in that: The aforementioned constraint loss on the physical mapping relationship of new energy output includes the constraint loss on the physical mapping relationship of wind power output. and the constraint loss of the physical mapping relationship of photovoltaic output Specifically: in, for t The wind power output value generated in real time; The theoretical output calculated based on wind speed and the turbine power curve; for t The photovoltaic output value generated at any time; The temperature coefficient of photovoltaic power. for t Ambient temperature at all times; Temperature under standard test conditions; This is the theoretical maximum output of the photovoltaic unit calculated based on the current light intensity.

9. The time-series simulation method for uncertain new energy power systems according to claim 6, characterized in that: When training the power output data simulation model, a phased training method is adopted, which includes a pre-training phase and an adversarial training phase. During the pre-training phase, the weighted sum of the output upper and lower limit constraint loss and the output change rate constraint loss is used as the loss term to train the bidirectional RNN separately. During the adversarial training phase, the discriminator parameters in the GAN are updated based on cross-entropy loss; the generator parameters are updated based on the generator loss function.

10. A time-series simulation system for new energy uncertain power systems based on GAN-RNN, according to the method of any one of claims 1-9, comprising a basic data acquisition module, a typical sample screening module, a power output data simulation model construction module, a power output data simulation model training module, a power output sequence generation module, and a system operation index calculation module, characterized in that: The basic data acquisition module acquires historical meteorological data for the whole year, as well as the corresponding new energy power output data and thermal power unit power output data at the same time, as the basic dataset; The typical sample selection module divides temperature intervals based on historical meteorological data throughout the year, calculates the proportion of days in each temperature interval to the total number of days in the year as the probability of occurrence of that temperature interval, calculates the scenario complexity based on the fluctuation range of new energy output within a single day, and selects typical monthly, typical weekly, and typical daily samples by combining the probability of occurrence of temperature intervals and scenario complexity. The power output data simulation model construction module constructs a power output data simulation model based on bidirectional RNN and GAN. The model has a built-in physical rule layer for embedding power system engineering constraints; and introduces physical constraint regularization terms into the generator loss function of GAN. The physical constraint regularization terms include output upper and lower limit constraint loss, output change rate constraint loss, and new energy output physical mapping relationship constraint loss. The power output data simulation model training module uses the basic dataset as training samples to train the power output data simulation model and obtain the trained model. The power output sequence generation module inputs meteorological data of typical monthly, typical weekly, and typical daily samples into the trained model to generate new energy power output sequences and thermal power unit power output sequences at the corresponding time scales. The system operation index calculation module calculates the corresponding system operation indexes based on the new energy output sequence and thermal power unit output sequence corresponding to typical months, typical weeks, and typical days. It also weights and superimposes the system operation indexes based on the probability of occurrence of the temperature range to which the typical month, typical week, and typical day belong, to obtain the final operation indexes.

Images