Intelligent prediction method and system for cascade hydropower generation capacity

Abstract

The application discloses a cascade hydropower generation capacity intelligent prediction method and system, belongs to the technical field of hydropower, and comprises the following steps: S1, four-step data preprocessing, obtaining complete'meteorological-hydrological-hydropower' sequences; S2, dividing a sample set; S3, constructing an intelligent interval prediction model coupled with GRU-DeepAR; S4, obtaining an initial interval prediction result; and S5, comprehensive error correction, obtaining a final prediction result. The application adopts the above cascade hydropower generation capacity intelligent prediction method, obtains high-robustness data through a four-step process, unifies probability prediction and non-negative constraints, and improves prediction reliability, accuracy and stability.

Description

Technical Field

[0001] This invention belongs to the field of hydropower technology, specifically relating to a method and system for intelligent prediction of the power generation capacity of cascade hydropower. Background Technology

[0002] Accurate prediction of hydropower generation capacity is a crucial foundation for flexible power grid dispatch, optimized resource allocation, efficient water resource utilization, and environmental protection. Hydropower generation capacity is influenced by a multitude of factors, including natural factors such as rainfall, runoff, and water levels, and human factors such as dispatch plans, operational efficiency, and load demand. The interplay of these factors makes it difficult for prediction models to fully capture the complex nonlinear and uncertain relationships between them, thus making the prediction of hydropower generation capacity exceptionally challenging.

[0003] However, existing technologies have the following shortcomings: raw meteorological-hydrological-hydropower data often contain missing values ​​and outliers, which directly affect the robustness of the model; traditional AI methods and probabilistic prediction frameworks are mostly simple stacking, failing to fully utilize the time series modeling capabilities and probabilistic generation advantages; as the forecast period increases, the coverage and width of interval predictions are prone to significant degradation, lacking effective post-correction methods; traditional methods are mostly single-value predictions, unable to provide confidence intervals, making it difficult to meet the risk management needs of scheduling decisions.

[0004] Therefore, a new method is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent prediction method for the power generation capacity of cascade hydropower. This method obtains highly robust data through a four-step process: extreme value and negative value removal, feature construction and standardization, K-nearest neighbor similar day interpolation, linear interpolation and physical constraint post-processing; and unifies probabilistic prediction and non-negative constraints to improve the reliability, accuracy and stability of prediction.

[0006] To achieve the above objectives, this invention provides an intelligent prediction method for the power generation capacity of cascade hydropower, comprising the following steps: S1. For missing values ​​and outliers in the original data, a four-step data preprocessing method is adopted to obtain the imputation results of missing values ​​and outliers, and to supplement them into a complete "meteorological-hydrological-hydropower" sequence; S2. Based on the "meteorological-hydrological-hydropower" sequence of S1, the data is split from the complete sequence according to a set ratio. The "meteorological-hydrological" related data of each sample is defined as the set of influencing factors, and the hydropower generation capacity value at the corresponding time is defined as the target value, so as to obtain the training sample set and the test sample set. S3. Based on the training sample set of S2, a gated recurrent unit and a deep autoregressive network are coupled. The gated recurrent unit is used to capture the dynamic dependence of the input data time series, and the deep autoregressive network is used to model the conditional probability distribution. The predicted mean and standard deviation are output to obtain the intelligent interval prediction model of hydropower generation capacity. S4. Input the set of influencing factors from the test sample set in S2 into the intelligent interval prediction model in S3 to obtain the corresponding initial interval prediction results of hydropower generation capacity. S5. Based on the prediction results of S4, a comprehensive error correction method is adopted to further reduce the deviation of the hydropower generation capacity prediction results and obtain the final hydropower generation capacity prediction results.

[0007] Preferably, S1 includes the following steps: S101. Remove extreme and negative values ​​from the original "meteorological-hydrological-hydropower" sequence: for inflow... Downflow Upstream water level Downstream water level and hydropower generation capacity First, the quantile threshold method is used to identify extreme values; denoted by variable... exist t The observed value at time is Given upper and lower quantiles, it is represented as: ; ; ; ; in, For variables Empirical quantile function; It is the lower quantile; High quantile; For variables In quantiles The specific value at the location; For variables In quantiles The specific value at the location; like If the observation is not found, it will be considered an anomaly and marked as missing; furthermore, for the variables of flow rate and hydropower generation capacity, if For water level variables, directly mark them as missing; Also marked as missing; The above operations generate a missing mask, calculated as follows: ; in, ; For the first t The moment, the first v Missing status markers for each variable; For the first t The moment, the first v The original observations of each variable; This refers to precipitation. S102. To capture the diurnal cyclical characteristics of "meteorological-hydrological-hydropower" data, time-derived variables are introduced: ; in, For the first The month corresponding to each moment; For the first The accumulated days of each time point; At the same time, calculate the flow difference. The formula is as follows: ; The final constructed feature vector is represented as follows: ; in, For temperature, For wind speed, Relative humidity, Solar radiation, This refers to daily precipitation. This refers to the soil moisture content in the root zone. This refers to the soil moisture content in the profile. It is a 10-dimensional real number space; The Z-score method is used to standardize different features, and the calculation method is as follows: ; in, and These are diagonal matrices representing the mean and standard deviation of the features in the training set, respectively. For the standardized first One sample; For the original first One sample; S103. Perform K-nearest neighbor similarity interpolation on outliers, for any variable with missing values. Utilizing its missing samples Construct a K-nearest neighbor model; Let the set of samples to be imputed be... For each Search for it In KThe samples with the smallest Euclidean distance form the nearest neighbors, denoted as: ; in, In order to be in At any given time, the values ​​filtered by Euclidean distance include A set of similar samples for each sample; For this set An index of similar samples; any , and The Euclidean distance is calculated as follows: ; in, For the target time The corresponding standardized sample vector; For the target time With sample The distance between them; The sample weight calculation method is defined as follows: ; in, For the target time With sample Weights between them; Let be a local constant, take The weights satisfy Furthermore, the closer the neighbor is, the greater its weight. ; Missing values The interpolation result is expressed as: ; in, For the target time The interpolated variable values; For the sample The corresponding original variable values; S104. Perform linear interpolation and physical constraint post-processing on the interpolated sequence. If still If isolated missing points always exist, then... Time missing point Perform bidirectional linear interpolation, using the following formula: ; in, For a moment The known values ​​of the variables; For a moment The known values ​​of the variables; Let the interpolated sequence be denoted as To ensure that the "meteorological-hydrological-hydropower" sequence satisfies the non-negative physical meaning, for all moments... Execution, the formula is as follows: ; The final complete "meteorology-hydrology-hydropower" sequence is as follows: .

[0008] Preferably, S3 includes the following steps: S301, Calculate using the gating mechanism of the gated loop unit. It constantly updates and resets the door's information values ​​and calculates... The external state information value of the hidden layer at any given time is calculated as follows: ; ; ; ; ; ; in, For element-wise multiplication; and These are the activation functions; for The external state of the hidden layer at all times; for The external state of the hidden layer at all times; for Candidate states of the hidden layer for "meteorological-hydrological-hydropower" time series data , To update the door; To reset the door; , This is the weight matrix; It is the bias vector; This is the input feature vector at the current time. for By updating the weight matrix at the gate time for By updating the weight matrix at the gate time To update the gate bias vector; S302. A deep autoregressive network framework is introduced, combined with a gated recurrent unit as a sequence encoder, to achieve unified modeling of probabilistic prediction and non-negativity constraints. The gated recurrent unit encoder is used to extract the temporal features of "meteorological-hydrological-hydropower" data, and the output layer of the deep autoregressive network is used to output Gaussian parameters and obtain the initial interval prediction results. Given a "meteorological-hydrological-hydropower" input sequence The gated recurrent unit updates the hidden state time-by-time and outputs the hidden state of the last time step as a high-dimensional representation of the global sequence representation: ; in, No. At that moment The hidden state output is for the previous Comprehensive encoding of the temporal information of the input sequence at each time point; For the first The input feature vector at each time step; Assume the target variable is hydropower generation capacity. Follows a Gaussian distribution: ; in, To predict the mean of the sequence, The standard deviation of the predicted sequence; So, The mean and standard deviation of the fully connected layer when mapped to Gaussian parameters are calculated as follows: ; ; in, , These are the weight matrix and bias term calculated from the mean, respectively; The standard deviation of the predicted distribution of hydropower generation capacity. , These are the weight matrix and bias term calculated from the standard deviation, respectively; The final range prediction results for hydropower generation capacity are as follows: ; in, These are the upper and lower bounds for interval prediction, respectively. These are the quantiles of the standard normal distribution; Introducing a constraint term to ensure the lower bound of the interval prediction is non-negative, the loss function with the constraint term is calculated as follows: ; in, The value of the loss function. This represents the actual observed value of hydropower generation capacity. To predict the mean of the distribution, The penalty weight is set to 0.1; These are non-negative constraint terms. For activation function, .

[0009] Preferably, S5 includes the following steps: S501. To make the coverage of the interval prediction results approach the target coverage, dynamically adjust... The value is represented as: ; in, , These represent the true coverage and target coverage of the sliding window, respectively. The window length in this paper is set to 30. Let be the learning rate, and take... ; and They are respectively The upper and lower bounds of the value are taken. , , For dynamic adjustment functions, Values ​​are restricted to a range middle; S502. The point prediction results obtained by using a long short-term memory network to couple a gated recurrent unit and a deep autoregressive network are corrected and expressed as follows: ; ; in, and They are respectively t The forgetting gates and hidden layers of the long short-term memory network; and They are respectively t The external state and memory units of the long short-term memory network at all times; For length is The residual history window serves as the input to the long short-term memory network; This represents the residual prediction value output by the Long Short-Term Memory network; The predicted hydropower generation capacity points after correction using the Long Short-Term Memory (LSTM) network are represented as follows: ; in, and They are respectively t Initial results of time-based hydropower generation capacity prediction and results after correction using the long short-term memory network; S503. Using the robust variance scaling method, the predicted standard deviation is adjusted based on the empirical standard deviation of the residuals. The calculation formula is as follows: ; ; in, and These represent the standard deviations of the hydropower generation capacity forecast results before and after adjustment. Adjust the standard deviation ratio. for t -30 o'clock t The empirical standard deviation of the time-based prediction results; and They are respectively The upper and lower bounds of the value are taken. , , For dynamic adjustment functions, Values ​​are restricted to a range middle; S504. The corrected upper and lower limits of the prediction interval are expressed as follows: ; ; in, This is the upper bound of the predicted range for hydropower generation capacity after correction. The lower bound is used to define the predicted range of hydropower generation capacity after correction. For the first The corrected predicted mean at each time point.

[0010] This invention also provides an intelligent prediction system for the power generation capacity of cascade hydropower, including: The data preprocessing module is used to address missing and outlier values ​​in the original data by employing a four-step data preprocessing method to obtain the imputation results of missing and outlier values, thus supplementing them into a complete "meteorological-hydrological-hydropower" sequence. The sample partitioning module, connected to the data preprocessing module, splits the data from the complete sequence according to a set ratio based on the "meteorology-hydrology-hydropower" sequence. It defines the "meteorology-hydrology" related data of each sample as a set of influencing factors and the hydropower generation capacity value at the corresponding time as the target value, thereby obtaining the training sample set and the test sample set. The model building module is connected to the sample partitioning module. Based on the training sample set, it couples a gated recurrent unit and a deep autoregressive network. The gated recurrent unit is used to capture the dynamic dependence of the input data time series, and the deep autoregressive network is used to model the conditional probability distribution. The predicted mean and standard deviation are output to obtain the intelligent interval prediction model of hydropower generation capacity. The initial prediction module, connected to the model building module, is used to input the set of influencing factors in the test sample set into the intelligent interval prediction model to obtain the corresponding initial interval prediction results of hydropower generation capacity. An error correction module, connected to the initial prediction module, uses a comprehensive error correction method based on the prediction results to further reduce the deviation of the hydropower generation capacity prediction results and obtain the final hydropower generation capacity prediction results.

[0011] Therefore, the present invention employs the above-mentioned intelligent prediction method for cascade hydropower generation capacity. Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) This invention obtains highly robust data through a four-step process: extreme value and negative value removal, feature construction and standardization, K-nearest neighbor similarity interpolation, linear interpolation and physical constraint post-processing. (2) This invention adopts a coupled GRU and DeepAR framework. GRU captures temporal dynamic dependencies, and DeepAR models the conditional probability distribution to achieve the unification of probability prediction and non-negativity constraints. (3) The present invention adopts a three-level correction method of dynamic confidence level adjustment, LSTM residual correction and robust variance scaling to improve the reliability, accuracy and stability of prediction; (4) This invention provides confidence intervals by outputting interval prediction results containing mean and standard deviation, thereby meeting the needs of scheduling decisions for risk quantification.

[0012] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0013] Figure 1 This is a flowchart of an embodiment of the intelligent prediction method for cascade hydropower generation capacity of the present invention; Figure 2 This is a comparison chart of hydropower generation capacity prediction results in an embodiment of the intelligent prediction method for cascade hydropower generation capacity of the present invention. Figure 2 In the text, (a) represents the comparison between the prediction results of the DeepAR model and the actual observed values; Figure 2 (b) in the figure represents the comparison between the prediction results of the GRU-DeepAR model and the actual observed values; Figure 2 (c) in the figure represents the comparison between the prediction results of the GRU-DeepAR-EC model and the actual observations. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0015] Example 1 like Figures 1-2 As shown, the intelligent prediction method for cascade hydropower generation capacity of the present invention includes the following steps: S1. To address missing and outlier values ​​in the original data, a four-step data preprocessing method is employed to obtain imputation results for missing and outlier values, supplementing them into a complete "meteorological-hydrological-hydropower" sequence; including the following steps: S101. Remove extreme and negative values ​​from the original "meteorological-hydrological-hydropower" sequence: for inflow... Downflow Upstream water level Downstream water level and hydropower generation capacity First, the quantile threshold method is used to identify extreme values; denoted by variable... exist t The observed value at time is Given upper and lower quantiles, it is represented as: ; ; ; ; in, For variables Empirical quantile function; It is the lower quantile; High quantile; For variables In quantiles The specific value at the location; For variables In quantiles The specific value at the location; like If the observation is not found, it will be considered an anomaly and marked as missing; furthermore, for the variables of flow rate and hydropower generation capacity, if For water level variables, directly mark them as missing; Also marked as missing; The above operations generate a missing mask, calculated as follows: ; in, ; For the first The moment, the first Missing status markers for each variable; For the first The moment, the first The original observations of each variable; This refers to precipitation. S102. Feature construction and standardization of raw data: In order to capture the diurnal cyclical characteristics of "meteorological-hydrological-hydropower" data, time-derived variables are introduced: ; in, For the first The month corresponding to each moment; For the first The accumulated days of each time point; At the same time, calculate the flow difference. The formula is as follows: ; The final constructed feature vector is represented as follows: ; in, For temperature, For wind speed, Relative humidity, Solar radiation, This refers to daily precipitation. This refers to the soil moisture content in the root zone. This refers to the soil moisture content in the profile. It is a 10-dimensional real number space; To avoid the influence of different units of measurement, the Z-score method is used to standardize different features. The calculation method is as follows: ; in, and These are diagonal matrices representing the mean and standard deviation of the features in the training set, respectively. For the standardized first One sample; For the original first One sample; S103. After identifying missing values ​​and the feature matrix, further processing based on K-nearest neighbor similarity interpolation is needed for outliers: for any variable with missing values... Utilizing its missing samples Construct a K-nearest neighbor model; Let the set of samples to be imputed be... For each Search for it In K The samples with the smallest Euclidean distance form the nearest neighbors, denoted as: ; in, In order to be in At any given time, the values ​​filtered by Euclidean distance include A set of similar samples for each sample; For this set An index of similar samples; any , and The Euclidean distance is calculated as follows: ; in, For the target time The corresponding standardized sample vector; For the target time With sample The distance between them; To avoid division by zero, a small constant is introduced. The sample weight calculation method is defined as follows: ; in, For the target time With sample Weights between them; Let be a local constant, take ; Weights satisfy Furthermore, the closer the neighbor is, the greater its weight. ; Missing values The interpolation result is expressed as: ; in, For the target time The interpolated variable values; For the sample The corresponding original variable values; S104. Perform linear interpolation and physical constraint post-processing on the interpolated sequence. If still If isolated missing points always exist, then... Momentary gaps Perform bidirectional linear interpolation, as shown in the following formula: ; in, For a moment The known values ​​of the variables; For a moment The known values ​​of the variables; Let the interpolated sequence be denoted as To ensure that the "meteorological-hydrological-hydropower" sequence satisfies the non-negative physical meaning, for all moments... Execution, the formula is as follows: ; The final complete "meteorology-hydrology-hydropower" sequence is as follows: ; S2. Based on the "meteorological-hydrological-hydropower" sequence of S1, the data is split from the complete sequence according to a set ratio. The "meteorological-hydrological" related data of each sample are defined as the set of influencing factors, and the hydropower generation capacity value at the corresponding time is defined as the target value, thus obtaining the training sample set. and test sample set ,in , They represent the first The set of influencing factors for each training sample, and the hydropower generation capacity value. , They represent the first The set of influencing factors for each test sample, and the hydropower generation capacity value. and These represent the number of training samples and the number of test samples, respectively. S3. Based on the training sample set of S2, a gated recurrent unit and a deep autoregressive network are coupled. The gated recurrent unit is used to capture the dynamic dependence of the input data time series, and the deep autoregressive network is used to model the conditional probability distribution, outputting the predicted mean and standard deviation, thus obtaining an intelligent interval prediction model for hydropower generation capacity; including the following steps: S301, Calculate using the gating mechanism of the gated loop unit. t It constantly updates and resets the door's information values ​​and calculates... t The external state information value of the hidden layer at any given time is calculated as follows: ; ; ; ; ; ; in, For element-wise multiplication; and These are the activation functions; for The external state of the hidden layer at all times; for The external state of the hidden layer at all times; for Candidate states of the hidden layer for "meteorological-hydrological-hydropower" time series data , To update the door; To reset the door; , This is the weight matrix; It is the bias vector; This is the input feature vector at the current time. for By updating the weight matrix at the gate time for By updating the weight matrix at the gate time To update the gate bias vector; Traditional intelligent prediction models such as S302 and gated cyclic units only output single-value predictions and cannot provide confidence intervals. Therefore, a deep autoregressive network framework is introduced, combined with a gated cyclic unit as a sequence encoder, to achieve unified modeling of probabilistic prediction and non-negativity constraints. The gated cyclic unit encoder is used to extract the temporal features of "meteorological-hydrological-hydropower" data, and the deep autoregressive network output layer is used to output Gaussian parameters and obtain initial interval prediction results.

[0016] Given a "meteorological-hydrological-hydropower" input sequence The gated recurrent unit updates the hidden state time-by-time and outputs the hidden state of the last time step as a high-dimensional representation of the global sequence representation: ; in, No. At that moment The hidden state output is for the previous Comprehensive encoding of the temporal information of the input sequence at each time point; For the first The input feature vector at each time step; Assume the target variable is hydropower generation capacity. Follows a Gaussian distribution: ; in, To predict the mean of the sequence, The standard deviation of the predicted sequence; So, The mean and standard deviation of the fully connected layer when mapped to Gaussian parameters are calculated as follows: ; ; in, , These are the weight matrix and bias term calculated from the mean, respectively; The standard deviation of the predicted distribution of hydropower generation capacity. Let be a very small constant; to prevent numerical instability, take . ; , These are the weight matrix and bias term calculated from the standard deviation, respectively; The final range prediction results for hydropower generation capacity are as follows: ; in, These are the upper and lower bounds for interval prediction, respectively. These are the quantiles of the standard normal distribution; On the other hand, considering that negative values ​​have no physical meaning in the process of predicting hydropower generation capacity, a constraint term is introduced to ensure that the lower limit of the interval prediction is non-negative. The loss function with the constraint term is then calculated as follows: ; in, The value of the loss function. This represents the actual observed value of hydropower generation capacity. To predict the mean of the distribution, The penalty weight is set to 0.1; These are non-negative constraint terms. For activation function, ; S4. Input the set of influencing factors from the test sample set in S2 into the intelligent interval prediction model in S3 to obtain the corresponding initial interval prediction results of hydropower generation capacity. S5. Based on the prediction results of S4, a comprehensive error correction method is used to further reduce the deviation of the hydropower generation capacity prediction results, and the final hydropower generation capacity prediction results are obtained; including the following steps: S501. To make the coverage of the interval prediction results approach the target coverage, dynamically adjust... The value, specifically, is obtained through the following method: ; in, , These represent the true coverage and target coverage of the sliding window, respectively. The window length in this paper is set to 30. Let be the learning rate, and take... ; and They are respectively The upper and lower bounds of the value, take =3.0, , For dynamic adjustment functions, Values ​​are restricted to a range In the meantime, avoid over-adjustment; S502. Using a long short-term memory network to model the residuals of the point prediction results obtained from the coupled model of the gated recurrent unit and the deep autoregressive network, i.e., the mean of the interval prediction results, the point prediction results are corrected: ; ; in, and They are respectively t The forgetting gates and hidden layers of the long short-term memory network; and They are respectively t The external state and memory units of the long short-term memory network at all times; Indicates length is The residual history window serves as the input to the long short-term memory network; This represents the residual prediction value of the output of the Long Short-Term Memory network; The predicted hydropower generation capacity points after correction using the Long Short-Term Memory (LSTM) network are represented as follows: ; in, and They are respectively t Initial results of time-based hydropower generation capacity prediction and results after correction using the long short-term memory network; S503. To prevent overconfidence or overconservatism in the forecast results, a robust variance scaling method is adopted. The forecast standard deviation is adjusted based on the empirical standard deviation of the residuals. The calculation formula is as follows: ; ; in, and These represent the standard deviations of the hydropower generation capacity forecast results before and after adjustment. Adjust the standard deviation ratio. fort -30 o'clock t The empirical standard deviation of the time-based prediction results; and They are respectively The upper and lower bounds of the value, take =2.0, , For dynamic adjustment functions, Values ​​are restricted to a range In the meantime, avoid over-adjustment; S504, after integrating the three-level correction method, we obtain t Interval prediction results after time correction Represented as: ; ; in, and These represent the upper and lower bounds of the corrected hydropower generation capacity prediction range, respectively. For the first The corrected predicted mean at each time point.

[0017] A comparative study was conducted using daily-scale data on power generation capacity and inflow at a hydropower station. Seven evaluation indicators, including RMSE, MAE, and CC, were used to compare the performance with DeepAR and GRU-DeepAR models. Experimental results show that the prediction error of this embodiment is smaller, the fit with the measured values ​​is higher, and it has good prediction performance, as shown in Table 1.

[0018] Table 1. Evaluation Indicators of the Prediction Model

[0019] Therefore, this invention adopts the above-mentioned intelligent prediction method for cascade hydropower generation capacity. This method obtains highly robust data through a four-step process: extreme value and negative value removal, feature construction and standardization, K-nearest neighbor similar day interpolation, linear interpolation and physical constraint post-processing; and unifies probabilistic prediction and non-negative constraints to improve the reliability, accuracy and stability of prediction.

[0020] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0021] 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 them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent prediction of the power generation capacity of cascade hydropower, characterized in that, Includes the following steps: S1. For missing values ​​and outliers in the original data, a four-step data preprocessing method is adopted to obtain the imputation results of missing values ​​and outliers, and to supplement them into a complete "meteorological-hydrological-hydropower" sequence; The four-step data preprocessing method includes: Extreme and negative values ​​are removed from the original "meteorological-hydrological-hydropower" sequence; time-derived variables are introduced to capture the daily-scale cyclical characteristics of the "meteorological-hydrological-hydropower" data; outliers are processed by K-nearest neighbor similar daily interpolation; and the interpolated sequence is post-processed with linear interpolation and physical constraints. S2. Based on the "meteorological-hydrological-hydropower" sequence of S1, the data is split from the complete sequence according to a set ratio. The "meteorological-hydrological" related data of each sample is defined as the set of influencing factors, and the hydropower generation capacity value at the corresponding time is defined as the target value, so as to obtain the training sample set and the test sample set. S3. Based on the training sample set of S2, a gated recurrent unit and a deep autoregressive network are coupled. The gated recurrent unit is used to capture the dynamic dependence of the input data time series, and the deep autoregressive network is used to model the conditional probability distribution. The predicted mean and standard deviation are output to obtain the intelligent interval prediction model of hydropower generation capacity. S4. Input the set of influencing factors from the test sample set in S2 into the intelligent interval prediction model in S3 to obtain the corresponding initial interval prediction results of hydropower generation capacity. S5. Based on the prediction results of S4, a comprehensive error correction method is used to further reduce the deviation of the hydropower generation capacity prediction results, and the final hydropower generation capacity prediction results are obtained; including the following sub-steps: S501. To make the coverage of the interval prediction results approach the target coverage, dynamically adjust... The value is represented as: ； in, , These represent the true coverage and target coverage of the sliding window, respectively. The window length in this paper is set to 30. Let be the learning rate, and take... ; and They are respectively The upper and lower bounds of the value are taken. =3.0, , For dynamic adjustment functions, Values ​​are restricted to a range middle; These are the quantile coefficients; S502. The point prediction results obtained by using a long short-term memory network to couple a gated recurrent unit and a deep autoregressive network are corrected and expressed as follows: ； ； in, and They are respectively The forgetting gates and hidden layers of the long short-term memory network; and They are respectively The external state and memory units of the long short-term memory network at all times; For length is The residual history window serves as the input to the long short-term memory network; This represents the residual prediction value output by the Long Short-Term Memory network; The predicted hydropower generation capacity points after correction using the Long Short-Term Memory (LSTM) network are represented as follows: ； in, and They are respectively Initial results of time-based hydropower generation capacity prediction and results after correction using the long short-term memory network; S503. Using the robust variance scaling method, the predicted standard deviation is adjusted based on the empirical standard deviation of the residuals. The calculation formula is as follows: ； ； in, and These represent the standard deviations of the hydropower generation capacity forecast results before and after adjustment. Adjust the standard deviation ratio. for t -30 o'clock t The empirical standard deviation of the time-based prediction results; and They are respectively The upper and lower bounds of the value are taken. , , For dynamic adjustment functions, Values ​​are restricted to a range middle; S504. The corrected upper and lower limits of the prediction interval are expressed as follows: ； ； in, This is the upper bound of the predicted range for hydropower generation capacity after correction. The lower bound is used to define the predicted range of hydropower generation capacity after correction. For the first The corrected predicted mean at each time point.

2. The intelligent prediction method for cascade hydropower generation capacity according to claim 1, characterized in that, S1 includes the following steps: S101. Remove extreme and negative values ​​from the original "meteorological-hydrological-hydropower" sequence: for inflow... Downflow Upstream water level Downstream water level and hydropower generation capacity First, the quantile threshold method is used to identify extreme values; denoted by variable... exist t The observed value at time is Given upper and lower quantiles, it is represented as: ； ； ； ； in, For variables Empirical quantile function; It is the lower quantile; High quantile; For variables In quantiles The specific value at the location; For variables In quantiles The specific value at the location; like If the observation is not found, it will be considered an anomaly and marked as missing; furthermore, for the variables of flow rate and hydropower generation capacity, if For water level variables, directly mark them as missing; Also marked as missing; The above operations generate a missing mask, calculated as follows: ； in, ; For the first t The moment, the first v Missing status markers for each variable; For the first t The moment, the first v The original observations of each variable; This refers to precipitation. S102. To capture the diurnal cyclical characteristics of "meteorological-hydrological-hydropower" data, time-derived variables are introduced: ； in, For the first The month corresponding to each moment; For the first The accumulated days of each time point; At the same time, calculate the flow difference. The formula is as follows: ； The final constructed feature vector is represented as follows: ； in, For temperature, For wind speed, Relative humidity, Solar radiation, Daily precipitation This refers to the soil moisture content in the root zone. This refers to the soil moisture content in the profile. It is a 10-dimensional real number space; The Z-score method is used to standardize different features, and the calculation method is as follows: ； in, and These are diagonal matrices representing the mean and standard deviation of the features in the training set, respectively. For the standardized first One sample; For the original first One sample; S103. Perform K-nearest neighbor similarity interpolation on outliers, for any variable with missing values. Utilizing its missing samples Construct a K-nearest neighbor model; Let the set of samples to be imputed be... For each Search for it In K The samples with the smallest Euclidean distance form the nearest neighbors, denoted as: ； in, In order to be in At any given time, the values ​​filtered by Euclidean distance include A set of similar samples for each sample; For this set An index of similar samples; any , and The Euclidean distance is calculated as follows: ； in, For the target time The corresponding standardized sample vector; For the target time With sample The distance between them; The sample weight calculation method is defined as follows: ； in, For the target time With sample Weights between them; Let be a local constant, take The weights satisfy Furthermore, the closer the neighbor is, the greater its weight. ; Missing values The interpolation result is expressed as: ； in, For the target time t The interpolated variable values; For the sample The corresponding original variable values; S104. Perform linear interpolation and physical constraint post-processing on the interpolated sequence. If still If isolated missing points always exist, then... Momentary gaps Perform bidirectional linear interpolation, as shown in the following formula: ； in, For a moment The known values ​​of the variables; For a moment The known values ​​of the variables; Let the interpolated sequence be denoted as To ensure the "meteorological-hydrological-hydropower" sequence satisfies non-negative physical meaning, for all moments... Execution, the formula is as follows: ； The final complete "meteorology-hydrology-hydropower" sequence is as follows: .

3. The intelligent prediction method for cascade hydropower generation capacity according to claim 1, characterized in that, S3 includes the following steps: S301, Calculate using the gating mechanism of the gated loop unit. It constantly updates and resets the door's information values ​​and calculates... t The external state information value of the hidden layer at any given time is calculated as follows: ； ； ； ； ； ； in, For element-wise multiplication; and These are the activation functions; for The external state of the hidden layer at all times; for The external state of the hidden layer at all times; for Candidate states of the hidden layer for "meteorological-hydrological-hydropower" time series data , To update the door; To reset the door; , This is the weight matrix; It is the bias vector; This is the input feature vector at the current time. for By updating the weight matrix at the gate time for By updating the weight matrix at the gate time To update the gate bias vector; S302. A deep autoregressive network framework is introduced, combined with a gated recurrent unit as a sequence encoder, to achieve unified modeling of probabilistic prediction and non-negativity constraints. The gated recurrent unit encoder is used to extract the temporal features of "meteorological-hydrological-hydropower" data, and the output layer of the deep autoregressive network is used to output Gaussian parameters and obtain the initial interval prediction results. Given a "meteorological-hydrological-hydropower" input sequence The gated recurrent unit updates the hidden state time-by-time and outputs the hidden state of the last time step as a high-dimensional representation of the global sequence representation: ； in, No. At that moment The hidden state output is for the previous Comprehensive encoding of the temporal information of the input sequence at each time point; For the first The input feature vector at each time step; Assume the target variable is hydropower generation capacity. Follows a Gaussian distribution: ； in, To predict the mean of the sequence, The standard deviation of the predicted sequence; So, The mean and standard deviation of the fully connected layer when mapped to Gaussian parameters are calculated as follows: ； ； in, , These are the weight matrix and bias term calculated from the mean, respectively; The standard deviation of the predicted distribution of hydropower generation capacity. , These are the weight matrix and bias term calculated from the standard deviation, respectively; The final range prediction results for hydropower generation capacity are as follows: ； in, These are the upper and lower bounds for interval prediction, respectively. These are the quantiles of the standard normal distribution; Introducing a constraint term to ensure the lower bound of the interval prediction is non-negative, the loss function with the constraint term is calculated as follows: ； in, The value of the loss function. This represents the actual observed value of hydropower generation capacity. To predict the mean of the distribution, The penalty weight is set to 0.1; These are non-negative constraint terms. For activation function, .

4. A smart prediction system for cascade hydropower generation capacity, applied to the smart prediction method for cascade hydropower generation capacity as described in any one of claims 1-3, characterized in that, include: The data preprocessing module is used to address missing and outlier values ​​in the original data by employing a four-step data preprocessing method to obtain the imputation results of missing and outlier values, thus supplementing them into a complete "meteorological-hydrological-hydropower" sequence. The sample partitioning module, connected to the data preprocessing module, splits the data from the complete sequence according to a set ratio based on the "meteorology-hydrology-hydropower" sequence. It defines the "meteorology-hydrology" related data of each sample as a set of influencing factors and the hydropower generation capacity value at the corresponding time as the target value, thereby obtaining the training sample set and the test sample set. The model building module is connected to the sample partitioning module. Based on the training sample set, it couples a gated recurrent unit and a deep autoregressive network. The gated recurrent unit is used to capture the dynamic dependence of the input data time series, and the deep autoregressive network is used to model the conditional probability distribution. The predicted mean and standard deviation are output to obtain the intelligent interval prediction model of hydropower generation capacity. The initial prediction module, connected to the model building module, is used to input the set of influencing factors in the test sample set into the intelligent interval prediction model to obtain the corresponding initial interval prediction results of hydropower generation capacity. An error correction module, connected to the initial prediction module, uses a comprehensive error correction method based on the prediction results to further reduce the deviation of the hydropower generation capacity prediction results and obtain the final hydropower generation capacity prediction results.

5. A computer device, characterized in that, include: A processor configured to be coupled to memory, read and execute instructions and / or program code in the memory to perform the method as described in any one of claims 1-3.

6. A computer-readable medium, characterized in that, The computer-readable medium stores computer program code that, when executed on a computer, causes the computer to perform the method as described in any one of claims 1-3.

Images