A method for forward-looking risk deduction and early warning for water, wind and light daily scheduling

Abstract

This invention relates to a forward-looking risk projection and early warning method for intraday dispatching of hydropower, wind power, and solar power, belonging to the field of power grid operation and control technology. First, based on an improved LSTM-3D-CNN model, it extracts the complex spatiotemporal features between the power system operating state and multi-scenario ultra-short-term wind and solar runoff predictions. The LSTM captures time dependencies, while the 3D-CNN extracts local spatiotemporal patterns through an adaptive convolutional window. Second, using a multi-output Gaussian mixture density network (GMDN) with the extracted features as conditional input, it fits the conditional probability distributions of four types of operational risks—wind curtailment, solar curtailment, hydropower curtailment, and load shedding—within the forward-looking dispatch window. Finally, it sets risk levels and early warning levels based on risk probabilities and dynamically adjusts generator output or start-up / shutdown strategies. This invention transforms the traditional "post-event response" into "pre-event early warning," significantly improving the dispatching safety and economy of hydropower, wind power, and solar power systems under uncertain environments.

Description

Technical Field

[0001] This invention belongs to the field of power grid operation and control technology, and specifically relates to a forward-looking risk simulation and early warning method for intraday scheduling of hydropower, wind power and solar power. Background Technology

[0002] Taking hydropower as the lead and forming a new pattern of development and operation of basin-wide clean energy bases through multi-energy complementarity of hydropower, wind power, and solar power is a key approach to achieving the "dual carbon target." However, traditional intraday scheduling methods are clearly insufficient in reliability and applicability when facing complex and variable hydropower, wind power, and solar power operating environments and high-dimensional hydropower, wind power, and solar power scenarios. Therefore, how to effectively simulate and predict the system operation risks brought about by uncertainties, achieve forward-looking risk warnings, adjust system operation modes, and ensure the safe and economical operation of the system has become a critical issue that urgently needs to be addressed.

[0003] Existing intraday scheduling methods for water, wind, and solar power do not consider the system operation risks caused by the ultra-short-term uncertainties of these factors, and therefore cannot take preventative measures in advance. Deep learning, on the other hand, eliminates the need for manually designed features and can automatically mine nonlinear relationships and potential patterns in data through multi-layered neural networks. This provides support for establishing a mapping relationship between predicted values ​​of water, wind, and solar power scenarios, system operating states, and uncertainties at future moments, thereby enabling virtual simulation of water, wind, and solar power uncertainties. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a forward-looking risk projection and early warning method for intraday scheduling of hydropower, wind power, and solar power. By fully mining the scheduling operation data of hydropower, wind power, and solar power through deep learning, the method extracts the correlation features between the forward-looking operation risks of the system and the real-time operation status under the uncertainty of runoff and wind and solar power output. Based on a hybrid density network, a probability distribution of system operation risks is established, which enables the effective projection of the system's operation risks such as wind curtailment, solar curtailment, hydropower curtailment, and load shedding within the intraday forward-looking scheduling window. Different risk levels and early warning levels are set based on the risk probability, thereby effectively guiding scheduling decisions to reduce the risks of power curtailment, hydropower curtailment, and load shedding.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power, comprising: S1, extracting complex spatiotemporal features between power system operating status (time series data) and multi-scenario ultra-short-term prediction information based on an improved LSTM-3D-CNN model, wherein the improved LSTM-3D-CNN model includes an LSTM module and a 3D-CNN module; extracting the time-dependent features of the power system operating status (time series data) through the LSTM module; and extracting local spatiotemporal features through the 3D-CNN module, wherein the convolution window size of the 3D-CNN module can be dynamically adjusted to automatically adapt to the optimal convolution window size based on the input data features; S2. A multi-output Gaussian mixture density network is used to fit the probability distribution of multiple types of operational risks of the power system, so as to realize the inference of uncertain risks of water, wind and solar power. The multi-output Gaussian mixture density network uses the features extracted by the improved LSTM-3D-CNN model as conditional input, fits the conditional probability distribution of four types of operational risks of the power system within the look-ahead dispatch window by maximizing the likelihood estimation, and outputs the probability density function of the four types of operational risks. S3. Based on the conditional probability distribution of the output of the multi-output Gaussian mixture density network, set different risk levels and early warning levels, and dynamically adjust the generator set output and start-stop status based on the early warning levels.

[0006] Furthermore, in step S1, the input of the improved LSTM-3D-CNN model includes multi-scenario ultra-short-term prediction data of runoff and wind and solar power output, and current and historical operating status data of the power system, and the input data is normalized.

[0007] Furthermore, in step S1, the temporal features output by the LSTM module are reconstructed into a three-dimensional data cube form to adapt to the input requirements of the 3D-CNN module.

[0008] Furthermore, in step S1, the 3D-CNN module adopts an adaptive window design, which uses the variance and autocorrelation of the input data to provide intelligent initialization for the convolution window size, wherein the convolution window size in the time dimension is initialized based on the autocorrelation of the data.

[0009] Furthermore, in step S1, a pooling layer is set after the 3D-CNN module to reduce the feature dimension and enhance the feature representation capability.

[0010] Furthermore, in step S1, the temporal features extracted by the LSTM module are fused with the local spatiotemporal features extracted by the 3D-CNN module, and the fused comprehensive features are output through a fully connected layer.

[0011] Furthermore, the sub-step S1 is as follows: S1.1 The input data includes multi-scenario ultra-short-term forecast data of runoff and wind and solar power output, and current and historical operating status data of the power system, and the input data is normalized; the specific method is as follows: ; ; in, This is a collection of ultra-short-term forecast data for multiple scenarios of power system runoff and wind and solar power output. , and , respectively, are the ultra-short-term forecast values ​​of hydropower station i, photovoltaic power station g, and wind farm w at time t+1 under scenario s; NT is the number of moments in the intraday rolling scheduling cycle; This is a collection of power system operating status data. Let t be the water level of hydropower station i at time t. Let be the outflow from hydropower station i at time t-1; The input data to the model is normalized as follows: ; in, The normalized dataset; , For data sets The mean and variance; S1.2. Combine the normalized input data into a multidimensional time series, and extract the long-term temporal dependency features of the multidimensional time series using the LSTM module; the specific method is as follows: Using LSTM modules to capture long-term temporal dependencies in input data, , Combined into a multidimensional time series dataset: ; Where X is a multidimensional time series set; The normalized ultra-short-term runoff prediction data for hydropower station i; This is the normalized ultra-short-term power prediction data for photovoltaic power plant g; This is the normalized ultra-short-term power prediction data for wind farm w; This provides the current and historical water level status data for hydropower station i. For the current and historical outflow status data of hydropower station i, the LSTM module status update method is as follows: ; ; ; ; ; in, Let be the input vector at time t; Let be the hidden state at time t; Let t represent the state of the memory cell at time t. It is the sigmoid activation function; Represents element-wise product; , , , , , , , , , , , This is the parameter matrix of the LSTM network; The temporal feature output is obtained after passing through the LSTM layer: ; Where H is the output set of the LSTM; Let i be the i-th eigenvector; S1.3. The temporal features output by the LSTM module are reconstructed into a three-dimensional data cube and input into the 3D-CNN module; the convolution window size is intelligently initialized using the variance and autocorrelation of the input data, and local spatiotemporal features are extracted through three-dimensional convolution and pooling operations; the specific method is as follows: Reconstruct the features output by the LSTM module into a 3D data cube form: ; in, The data consists of 3D data after feature reconstruction; C represents the feature channels; M and W represent spatial dimensions. The length of the subsequence window; An adaptive mechanism is introduced, and the window size is determined based on an automatic learning method using a convolutional neural network, allowing... , , The control parameters for adaptive window size are expressed as: ; ; ; in, Ensure the window size is positive; clip is the truncation function; M, W, and Dc are the 3D convolution kernel sizes; , , , , , These are the upper and lower limits of the window size, which are set according to the data characteristics; Use data variance and autocorrelation to provide intelligent initialization for window size: ; ; in, , The initial window size M and W are control parameters; This represents the average variance of the data in the horizontal dimension; It is the overall mean of the variance of all locations or samples in the spatial dimension x; Initialize the window size for the time dimension using the autocorrelation of the data: ; in, For time step The autocorrelation coefficient under the following conditions; The data value at time t; This represents the data mean. The appropriate time window size is determined by first decaying the autocorrelation coefficient to a certain threshold; the time required to decay to the threshold is then used as the threshold value. As an initial estimate: ; in, Control parameters for the initial window size Dc; The formula for calculating 3D convolution is as follows: ; in, For output; These are the parameters for the 3D convolution kernel. For bias terms; Applying the ReLU activation function to the output y above, we obtain the convolution output: ; Further use pooling layers to reduce dimensionality and enhance feature representation: ; in, This represents the 3D max pooling operation.

[0012] S1.4. The temporal features extracted by the LSTM module are fused with the local spatiotemporal features extracted by the 3D-CNN module, and the fused comprehensive features are output through a fully connected layer. The specific method is as follows: Fusing features extracted by LSTM and 3D-CNN: ; Output after fully connected layer: ; In the formula, F is the feature fusion vector set; This represents a matrix concatenation operation; This is a matrix flattening operation; O represents the feature fusion output. , These are the weights and biases for the fully connected layer.

[0013] Furthermore, in step S2, the loss function of the multi-output Gaussian mixture density network adopts the negative log-likelihood function, and the network parameters are optimized through backpropagation. The loss function is as follows: ; Where T is the total number of samples in the training data; K is the number of Gaussian distributions in the Gaussian mixture model; The weight parameters are for the k-th Gaussian distribution; It follows a Gaussian distribution; , Let be the mean and standard deviation of the k-th Gaussian distribution.

[0014] Furthermore, in step S3, the risk level classification rule is as follows: if the probability of wind curtailment, solar curtailment, hydropower curtailment, or load loss exceeding a preset threshold is less than 0.4, then the risk level of the corresponding risk type is 1; if the probability is greater than or equal to 0.4, then the risk level is 2.

[0015] Furthermore, in step S3, when the risk level of any risk type is 1, the early warning level 1 measures are implemented: for wind curtailment risk or solar curtailment risk, reduce the output of generator units; for water curtailment risk, reduce the outflow from the reservoir; for load shedding risk, increase the output of generator units.

[0016] Furthermore, in step S3, when the risk of water wastage and the risk of load loss are both at risk level 2, the warning level 2 measures are implemented: an additional generator set is started.

[0017] Furthermore, in step S3, the adjustment amount of generator output and reservoir outflow under early warning level 1 are calculated according to a preset formula based on the risk simulation results.

[0018] A forward-looking risk simulation and early warning system for intraday scheduling of water, wind, and solar power includes: The data preprocessing module is used to acquire and normalize multi-scenario ultra-short-term forecast data of runoff and wind and solar power output, as well as power system operation status data; The spatiotemporal feature extraction module is used to extract complex spatiotemporal features using the improved LSTM-3D-CNN model; The risk simulation module is used to fit the conditional probability distribution of four types of operational risks—wind curtailment, solar curtailment, hydropower curtailment, and load shedding—through the multi-output Gaussian mixture density network. The early warning decision module is used to set risk levels and early warning levels according to the conditional probability distribution, and generate generator set scheduling and adjustment instructions.

[0019] A computer device includes one or more processors, on which one or more executable programs are stored. When the one or more executable programs are executed by the one or more processors, they are used to implement the forward-looking risk simulation and early warning method for intraday scheduling of water, wind and solar power.

[0020] A storage medium storing one or more executable programs, which, when executed, are used to implement the aforementioned forward-looking risk simulation and early warning method for intraday scheduling of water, wind, and solar power.

[0021] The present invention can achieve the following beneficial effects: 1) Traditional scheduling models rely on deterministic prediction and linear correlation assumptions, making it difficult to capture the complex coupling relationship between wind and solar power output and grid status. This invention proposes an improved LSTM-3D-CNN model. Through the design of a 3D-CNN network using LSTM and an adaptive window, it can fully mine the scheduling operation data of water, wind and solar power, and achieve adaptive extraction of temporal features and local complex features. Furthermore, through feature fusion, it extracts complex correlation features between multi-dimensional data such as runoff, the system's forward-looking operation risk under the uncertainty of wind and solar power output, real-time operation status, and water, wind and solar operation scenarios, which significantly improves the modeling accuracy under complex operating conditions.

[0022] 2) Traditional dispatching models mainly rely on post-event response, that is, adjusting dispatching decisions to minimize risks after they occur. This invention proposes a forward-looking prediction method for system operation risks based on hybrid density networks. By inputting real-time and historical system operation status, wind and solar ultra-short-term forecasts, and other data, it can accurately predict the risks of power curtailment, water curtailment, and load shedding within the system's intraday forward-looking dispatching window. This transforms the traditional "post-event response" model into a "pre-event warning" model, which is beneficial for the system to cope with complex operating environments and uncertainties of new energy sources.

[0023] 3) This invention proposes a method for quantifying system operation risk level and early warning level, and formulates scheduling strategy adjustment schemes under different risk levels, thereby effectively reducing system operation risk. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the forward-looking risk simulation and early warning process for intraday scheduling of water, wind and solar power according to an embodiment of the present invention; Figure 2 This is a system topology diagram of an embodiment of the present invention; Figure 3 This is a graph of the loss function during the training process in an embodiment of the present invention; Figure 4 This is a comparison diagram between the probability distribution derived from the embodiments of the present invention and the actual probability distribution; Figure 5 This is a probability comparison chart of uncertainty risks before and after the forward-looking early warning in an embodiment of the present invention. Detailed Implementation

[0025] Preferred solutions include Figures 1 to 5 As shown, a forward-looking risk simulation and early warning method for intraday scheduling of water, wind, and solar power is presented, with the overall process as follows: Figure 1 As shown.

[0026] The system topology diagram corresponding to the embodiment is as follows: Figure 2 As shown, the system includes four cascade hydropower stations, five photovoltaic power stations, and four wind farms.

[0027] A forward-looking risk simulation and early warning method for intraday scheduling of water, wind, and solar power is implemented for this system. The method includes the following steps: 1) Based on the improved LSTM-3D-CNN model, the complex spatiotemporal features between system operation status and multi-scenario ultra-short-term prediction information are extracted. The LSTM module can effectively extract the time dependence features of time series data, and the 3D-CNN module can dynamically adjust the size of the convolution window. The 3D-CNN network can automatically adapt to the optimal convolution window size according to the features of the input data, thereby enhancing the generalization and accuracy of the model.

[0028] The forward operational risks of a system are closely related to its current operating status and the predicted information of ultra-short-term wind and solar power output and runoff. By effectively extracting data features between the system's operating status and ultra-short-term prediction information across multiple scenarios, the correlation between forward-looking risk projection and known information can be better realized, improving the effectiveness and accuracy of forward-looking risk projection. Long Short-Term Memory (LSTM) networks can effectively process time-dependent sequence data. Through memory and forgetting mechanisms, they can capture long-term and short-term features in time series, making them particularly suitable for the time-varying nature of watershed inflow and wind and solar power output. 3D Convolutional Neural Networks (3D-CNNs) can effectively extract local features through convolution operations, making them especially suitable for identifying local patterns or regularities from high-dimensional, complex water, wind, and solar power scenarios and system states. An improved LSTM-3D-CNN model can extract complex spatiotemporal features from data. The 3D-CNN network, which can dynamically adjust the convolution window size, can automatically adapt to the optimal convolution window size based on the input data features, enhancing the model's generalization and accuracy. The steps for building the improved LSTM-3D-CNN model are as follows: S1.1 Determine the model input information The model's input consists of known data from the system's intraday forward scheduling process, including: multi-scenario ultra-short-term forecast data for runoff and wind and solar power output, and the system's current and historical operating status, as shown below:

[0029]

[0030] in, A collection of ultra-short-term forecast data for multiple scenarios of system runoff and wind and solar power output. , and , respectively, are the ultra-short-term forecast values ​​of hydropower station i, photovoltaic power station g, and wind farm w at time t+1 under scenario s; NT is the number of moments in the intraday rolling scheduling cycle; This is a collection of system operating status data. Let t be the water level of hydropower station i at time t. Let t be the outflow from hydropower station i at time t-1.

[0031] The input data to the model is normalized as follows:

[0032] in, The normalized dataset; , For data sets The mean and variance of.

[0033] S1.2 Extracting Time-Series Features of Data Using LSTM networks to capture long-term temporal dependencies in input data, , Combined into a multidimensional time series dataset:

[0034] Where X is a multidimensional time series set; The normalized ultra-short-term runoff prediction data for hydropower station i; This is the normalized ultra-short-term power prediction data for photovoltaic power plant g; This is the normalized ultra-short-term power prediction data for wind farm w; This provides the current and historical water level status data for hydropower station i. This provides the current and historical outflow status data for hydropower station i. The LSTM status update method is as follows:

[0035]

[0036]

[0037]

[0038]

[0039] in, Let be the input vector at time t; Let be the hidden state at time t; Let t represent the state of the memory cell at time t. It is the sigmoid activation function; Represents element-wise product; , , , , , , , , , , , This is the parameter matrix of the LSTM network.

[0040] The temporal feature output is obtained after passing through the LSTM layer:

[0041] Where H is the output set of the LSTM; Let be the i-th eigenvector. S1.3 New Information Input The information input process requires filtering the input data through a weight matrix and then multiplying it with an activation matrix to obtain the information input to the memory unit.

[0042]

[0043]

[0044] In the formula, and For bias, , , and This is the weight matrix. Let be the candidate value for the new information, and be the activation matrix of the input gate. S1.3 Local Feature Extraction Compared to two-dimensional convolutional networks, which can only extract features from planar data structures, three-dimensional convolutional networks (3D-CNNs) can process three-dimensional tensors by adding dimensions, enabling the simultaneous learning of temporal and spatially relevant features. 3D-CNN networks using adaptive window designs further mine the spatial-temporal local patterns of the data. To suit the input of 3D-CNNs, the features output by LSTM are reconstructed into a 3D data cube form.

[0045] in, The data consists of 3D data after feature reconstruction; C represents the feature channels; M and W represent spatial dimensions. This represents the length of the subsequence window.

[0046] An adaptive mechanism is introduced, and the window size is determined based on an automatic learning method using a convolutional neural network, allowing... , , The control parameter for adaptive window size (the initial value can be estimated from statistical data) is specifically expressed as follows:

[0047]

[0048]

[0049] in, Ensure the window size is positive; `clip` is a truncation function that limits the window size to a reasonable range; M, W, D c This refers to the 3D convolution kernel size; , , , , , These are the upper and lower limits of the window size, which are set according to the data characteristics.

[0050] Use data variance and autocorrelation to provide intelligent initialization for window size:

[0051]

[0052] in, , The initial window size M and W are control parameters; This represents the average variance of the data in the horizontal dimension; It is the overall mean of the variance of all locations or samples in the spatial dimension x.

[0053] Initialize the window size for the time dimension using the autocorrelation of the data:

[0054] in, For time step The autocorrelation coefficient under the following conditions; The data value at time t; This represents the data mean.

[0055] The appropriate time window size is determined by first assessing the autocorrelation coefficient's decay to a certain threshold. The time window for decay to reach the threshold is then used as the threshold value. As an initial estimate:

[0056] in, The control parameters for the initial window size Dc.

[0057] The formula for calculating 3D convolution is as follows:

[0058] in, For output; These are the parameters for the 3D convolution kernel. This is a bias term.

[0059] Applying the ReLU activation function to the output y above, we obtain the convolution output:

[0060] Further use pooling layers to reduce dimensionality and enhance feature representation:

[0061] in, This represents the 3D max pooling operation.

[0062] S1.4 Feature Fusion Fusing features extracted by LSTM and 3D-CNN:

[0063] Output after fully connected layer:

[0064] In the formula, F is the feature fusion vector set; This represents a matrix concatenation operation; This is a matrix flattening operation; O represents the feature fusion output. , These are the weights and biases for the fully connected layer.

[0065] 2) The multi-output Gaussian mixture density network (GMDN) is used to fit the probability distribution of multiple types of operational risks of the system, so as to realize the inference of uncertain risks of water, wind and solar power. The multi-output GMDN network uses the features extracted by the improved LSTM-3D-CNN model as conditional input, fits the conditional probability distribution of four types of operational risks of wind curtailment, solar curtailment, water curtailment and load loss of the system within the look-ahead window by maximizing the likelihood estimation, and outputs the probability density function of the four types of operational risks.

[0066] Gaussian Mixture Density Network (GMDN) uses a neural network to map the relationship between input data and the parameters of a Gaussian mixture distribution, thereby achieving a fit to the conditional probability distribution. Traditional GMDN networks are insufficient in capturing the complex spatiotemporal variations of data. An improved LSTM-3D-CNN network is used to extract complex temporal and local features, further enhancing the fitting performance of the GMDN. The feature O output from the improved LSTM-3D-CNN is input into the GMDN to obtain the probability density functions for various risks. A loss function L is constructed using maximum likelihood estimation for backpropagation optimization of the network. The loss function is as follows:

[0067] Where T is the total number of samples in the training data; K is the number of Gaussian distributions in the Gaussian mixture model; The weight parameters are for the k-th Gaussian distribution; It follows a Gaussian distribution; , Let be the mean and standard deviation of the k-th Gaussian distribution.

[0068] The curve of the loss function during the training of a hybrid density network is shown below. Figure 3As shown, in the initial stage, the model's loss is high, indicating that the model is unable to effectively capture the characteristics of these uncertain risks related to water, scenery, and landscape in the early stages, resulting in a poor fit. After the first few dozen iterations, both the training loss and the test loss decrease rapidly. This indicates that the model gradually captures the effective features of the system state and the water, scenery, and landscape scene information, and is able to learn the probability distribution characteristics of various risks. In the later stages of training, both the training loss and the test loss stabilize at small values, indicating that the model can effectively extrapolate various risks.

[0069] The comparison between the probability distribution of system operational risk derived from the hybrid density network and the actual probability distribution is as follows: Figure 4 As shown in the figure, the comparison results show that the calculated probability distributions of wind curtailment, solar curtailment, hydropower curtailment, and load shedding are close to the actual probability distributions, proving the effectiveness of the proposed method.

[0070] 3) Based on the conditional probability distribution output by the uncertainty risk deduction model, set different risk levels and early warning levels, and dynamically adjust the unit output and start-up / shutdown status based on the early warning level.

[0071] Different risk levels and warning levels are set using the probability distribution output from the uncertainty risk extrapolation model to adjust unit output and start-up / shutdown status. If the probability of power curtailment, water curtailment, and load shedding exceeding a certain threshold is less than 0.4, the risk level is set to 1; if it is greater than 0.4, the risk level is set to 2. The specific probability calculation formula is as follows:

[0072]

[0073]

[0074] in, Let be the amount of electricity discarded at section o at time t+1; Let be the amount of water discharged by hydropower station i at time t+1; Let be the amount of load loss at section o at time t+1; , and The thresholds for power wastage, water wastage, and load shedding are set to 0.1 pu, 0.2 pu, and 0.01 pu, respectively. and These are the probability distributions of power abandonment and load loss within section o at time t+1, respectively. Let be the probability distribution of the amount of water discarded by hydropower station i within section o at time t+1.

[0075] It can be obtained by convolving the probability distributions of wind and solar power curtailment within section o:

[0076] , These represent the probability distributions of wind and solar power curtailment within section o at time t.

[0077] The definitions and corresponding measures for different risk levels are shown in Table 1 below: Table 1. Definitions and Corresponding Measures for Warning Levels under Different Risk Levels

[0078] Under warning level 1, the formulas for calculating the increase or decrease in unit output due to power curtailment and load loss, and the decrease in outflow due to water curtailment are as follows:

[0079]

[0080] in, The increased output of unit n of hydropower station i within section o at time t+1 due to water curtailment, power curtailment, or load loss; The outflow of water within section o is the reduced outflow due to water abandonment within the reservoir.

[0081] Based on the calculated risk level and warning level, the scheduling decision-making method is adjusted, and the operational risks of the system before and after the adjustment are compared, such as... Figure 5 As shown, after considering forward-looking warnings and adjusting dispatch decisions, the expected values ​​for power curtailment, hydropower curtailment, and load shedding are 17.06MW, 59.95m³ / s, and 4.13MW, respectively, which are significantly lower than the 82.50MW, 100.43m³ / s, and 1.54MW without forward-looking warnings. This demonstrates that the proposed method can effectively reduce the uncertainty risk of multi-energy complementary systems of hydropower, wind power, and solar power.

[0082] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power, characterized in that, Includes the following steps: S1. Based on an improved LSTM-3D-CNN model, complex spatiotemporal features between the operating state of the power system and ultra-short-term prediction information for multiple scenarios are extracted. The improved LSTM-3D-CNN model includes an LSTM module and a 3D-CNN module. The LSTM module extracts the time-dependent features of the operating state of the power system. The 3D-CNN module extracts local spatiotemporal features. The convolution window size of the 3D-CNN module can be dynamically adjusted to automatically adapt to the optimal convolution window size based on the input data features. S2. A multi-output Gaussian mixture density network is used to fit the probability distribution of multiple types of operational risks of the power system, so as to realize the inference of uncertain risks of water, wind and solar power. The multi-output Gaussian mixture density network uses the features extracted by the improved LSTM-3D-CNN model as conditional input, fits the conditional probability distribution of four types of operational risks of the power system within the look-ahead dispatch window by maximizing the likelihood estimation, and outputs the probability density function of the four types of operational risks. S3. Based on the conditional probability distribution of the output of the multi-output Gaussian mixture density network, set different risk levels and early warning levels, and dynamically adjust the generator set output and start-stop status based on the early warning levels.

2. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S1, the input of the improved LSTM-3D-CNN model includes multi-scenario ultra-short-term prediction data of runoff and wind and solar power output, and current and historical operating status data of the power system, and the input data is normalized.

3. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S1, the temporal features output by the LSTM module are reconstructed into a three-dimensional data cube to adapt to the input requirements of the 3D-CNN module.

4. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S1, the 3D-CNN module adopts an adaptive window design, which uses the variance and autocorrelation of the input data to provide intelligent initialization for the convolution window size, wherein the convolution window size in the time dimension is initialized based on the autocorrelation of the data.

5. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S1, a pooling layer is set after the 3D-CNN module to reduce the feature dimension and enhance the feature representation capability.

6. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S1, the temporal features extracted by the LSTM module are fused with the local spatiotemporal features extracted by the 3D-CNN module, and the fused comprehensive features are output through a fully connected layer.

7. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, The sub-step of step S1 is as follows: S1.1 The input data includes multi-scenario ultra-short-term forecast data of runoff and wind and solar power output, and current and historical operating status data of the power system, and the input data is normalized; the specific method is as follows: ； ； in, This is a collection of ultra-short-term forecast data for multiple scenarios of power system runoff and wind and solar power output. , and , respectively, are the ultra-short-term forecast values ​​of hydropower station i, photovoltaic power station g, and wind farm w at time t+1 under scenario s; NT is the number of moments in the intraday rolling scheduling cycle; This is a collection of power system operating status data. Let t be the water level of hydropower station i at time t. Let be the outflow from hydropower station i at time t-1; The input data to the model is normalized as follows: ； in, The normalized dataset; , For data sets The mean and variance; S1.

2. Combine the normalized input data into a multidimensional time series, and extract the long-term temporal dependency features of the multidimensional time series using the LSTM module; the specific method is as follows: Using LSTM modules to capture long-term temporal dependencies in input data, , Combined into a multidimensional time series dataset: ； Where X is a multidimensional time series set; The normalized ultra-short-term runoff prediction data for hydropower station i; This is the normalized ultra-short-term power prediction data for photovoltaic power plant g; This is the normalized ultra-short-term power prediction data for wind farm w; This provides the current and historical water level status data for hydropower station i. For the current and historical outflow status data of hydropower station i, the LSTM module status update method is as follows: ； ； ； ； ； in, Let be the input vector at time t; Let be the hidden state at time t; Let t represent the state of the memory cell at time t. It is the sigmoid activation function; Represents element-wise product; , , , , , , , , , , , This is the parameter matrix of the LSTM network; The temporal feature output is obtained after passing through the LSTM layer: ； Where H is the output set of the LSTM; Let i be the i-th eigenvector; S1.

3. The temporal features output by the LSTM module are reconstructed into a three-dimensional data cube and input into the 3D-CNN module; the convolution window size is intelligently initialized using the variance and autocorrelation of the input data, and local spatiotemporal features are extracted through three-dimensional convolution and pooling operations; the specific method is as follows: Reconstruct the features output by the LSTM module into a 3D data cube form: ； in, The data consists of 3D data after feature reconstruction; C represents the feature channels; M and W represent spatial dimensions. The length of the subsequence window; An adaptive mechanism is introduced, and the window size is determined based on an automatic learning method using a convolutional neural network, allowing... , , The control parameters for adaptive window size are expressed as: ； ； ； in, Ensure the window size is positive; clip is the truncation function; M, W, and Dc are the 3D convolution kernel sizes; , , , , , These are the upper and lower limits of the window size, which are set according to the data characteristics; Use data variance and autocorrelation to provide intelligent initialization for window size: ； ； in, , The initial window size M and W are control parameters; This represents the average variance of the data in the horizontal dimension; It is the overall mean of the variance of all locations or samples in the spatial dimension x; Initialize the window size for the time dimension using the autocorrelation of the data: ； in, For time steps The autocorrelation coefficient under the following conditions; The data value at time t; This represents the data mean. The appropriate time window size is determined by first decaying the autocorrelation coefficient to a certain threshold; the time required to decay to the threshold is then used as the threshold value. As an initial estimate: ； in, Control parameters for the initial window size Dc; The formula for calculating 3D convolution is as follows: ； in, For output; These are the parameters for the 3D convolution kernel. For bias terms; Applying the ReLU activation function to the output y above, we obtain the convolution output: ； Further use pooling layers to reduce dimensionality and enhance feature representation: ； in, This represents a 3D max pooling operation; S1.

4. The temporal features extracted by the LSTM module are fused with the local spatiotemporal features extracted by the 3D-CNN module, and the fused comprehensive features are output through a fully connected layer. The specific method is as follows: Fusing features extracted by LSTM and 3D-CNN: ； Output after fully connected layer: ； In the formula, F is the feature fusion vector set; This represents a matrix concatenation operation; This is a matrix flattening operation; O represents the feature fusion output. , These are the weights and biases for the fully connected layer.

8. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S2, the loss function of the multi-output Gaussian mixture density network adopts the negative log-likelihood function, and the network parameters are optimized through backpropagation. The loss function is as follows: ； Where T is the total number of samples in the training data; K is the number of Gaussian distributions in the Gaussian mixture model; The weight parameters are for the k-th Gaussian distribution; It follows a Gaussian distribution; , Let be the mean and standard deviation of the k-th Gaussian distribution.

9. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S3, the risk level classification rule is as follows: if the probability of wind curtailment, solar curtailment, hydropower curtailment, or load loss exceeding a preset threshold is less than 0.4, then the risk level of the corresponding risk type is 1; if the probability is greater than or equal to 0.4, then the risk level is 2.

10. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S3, when the risk level of any risk type is 1, the early warning level 1 measures are implemented: for wind curtailment risk or solar curtailment risk, reduce the output of generator units; for water curtailment risk, reduce the outflow from the reservoir; for load shedding risk, increase the output of generator units.

11. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S3, when the risk of water wastage and the risk of load loss are both at risk level 2, the warning level 2 measures are implemented: an additional generator set is started.

12. The forward-looking risk projection and early warning method for intraday scheduling of water, wind, and solar power according to claim 1, characterized in that, In step S3, the adjustment amount of generator output and reservoir outflow under early warning level 1 are calculated according to the preset formula based on the risk simulation results.

13. A forward-looking risk simulation and early warning system for intraday scheduling of water, wind, and solar power, implementing the method as described in any one of claims 1 to 11, characterized in that, include: The data preprocessing module is used to acquire and normalize multi-scenario ultra-short-term forecast data of runoff and wind and solar power output, as well as power system operation status data; The spatiotemporal feature extraction module is used to extract complex spatiotemporal features using the improved LSTM-3D-CNN model; The risk simulation module is used to fit the conditional probability distribution of four types of operational risks—wind curtailment, solar curtailment, hydropower curtailment, and load shedding—through the multi-output Gaussian mixture density network. The early warning decision module is used to set risk levels and early warning levels according to the conditional probability distribution, and generate generator set scheduling and adjustment instructions.

14. A computer device, characterized in that: It includes one or more processors, on which one or more executable programs are stored. When the one or more executable programs are executed by the one or more processors, they are used to implement the forward-looking risk simulation and early warning method for intraday scheduling of water, wind and solar power as described in any one of claims 1-13.

15. A storage medium, characterized in that: It stores one or more executable programs, which, when executed, are used to implement the forward-looking risk simulation and early warning method for intraday scheduling of water, wind and solar power as described in any one of claims 1-13.

Images