A multi-dimensional collaborative iterative Bayesian deep learning water level prediction method and system
Through the Bayesian deep learning water level prediction method with multi-dimensional collaborative iteration, using Bayesian deep learning model and dynamic Bayesian network, the problem of insufficient accuracy of future water level prediction in the existing technology is solved, and higher prediction accuracy is achieved and reliability, adapt to complex water conservancy scenarios.
Patent Information
- Application Number
- CN202411259787.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-10
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2044-09-10
AI Technical Summary
The existing technology has low accuracy in future water level prediction and cannot effectively predict the changing trend of water level, resulting in challenges in flood control and disaster reduction and water conservancy scheduling.
Using the Bayesian deep learning water level prediction method with multi-dimensional collaborative iteration, the Bayesian deep learning model is constructed by acquiring and processing water conservancy big data, combining the basic residual neural network and dynamic Bayesian network to perform iterative prediction, and using Bayesian The posterior probability and KL divergence optimization model parameters are used to achieve the accuracy and reliability of water level prediction.
It improves the accuracy and reliability of water level prediction, provides confidence information of the prediction results, enables the model to be continuously updated and adjusted, adapts to the dynamic changes of the water conservancy system, and enhances the stability and adaptability of the prediction.
Smart Images
Figure CN118779619B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of water level prediction, and in particular to a multi-dimensional collaborative iterative Bayesian deep learning water level prediction method and system. Background Art
[0002] At present, the prediction of river water level and flow is related to major livelihood issues such as flood prevention and disaster reduction, and water conservancy scheduling. Data-driven water level and flow prediction generally only focuses on its own changing rules, and often uses autoregression methods for prediction. The typical single cycle of the autoregression method only uses the previous water conservancy information to predict the water level information on the next side. The data information affecting the water level elements is not fully mined, and its accuracy and reliability are not strong.
[0003] The patent with application number CN201910054629.7 discloses a method for predicting and warning the water level of a sponge city river, including obtaining the water level value of the river, the weather characteristic value and / or the river dam characteristic value within a unit time; drawing a continuous fluctuation graph between the water level value of the inner river and time; according to the continuous fluctuation graph and the weather characteristic value and / or the river dam characteristic value, determining whether the current water level of the inner river reaches the warning standard, if it reaches the warning standard, sending a warning signal to the data processing terminal to drain the inner river in advance. In the embodiment of the present invention, the water level value of the inner river, the weather characteristic value and the river dam characteristic value are obtained and drawn into a continuous fluctuation graph respectively, and whether the warning standard is reached according to the trend of the fluctuation graph, and then a warning signal is sent, so as to effectively predict and warn the rising trend of the water level in the future period of time, thereby avoiding the occurrence of urban waterlogging due to the water level rising too fast and the drainage too slow. However, the existing technology can only predict the rising or falling trend of the future water level, and cannot accurately predict the future water level. Summary of the invention
[0004] In order to solve the deficiency of the prior art that the future water level cannot be predicted more accurately, the present application discloses a multi-dimensional collaborative iterative Bayesian deep learning water level prediction method which can predict the future water level with more reliable and accurate accuracy.
[0005] The present invention discloses a multi-dimensional collaborative iterative Bayesian deep learning water level prediction method, which includes S1: acquiring water conservancy information data several hours before the prediction time through a water conservancy big data acquisition module, S2: processing the water conservancy information data into standardized data in a water conservancy time series through a water conservancy big data processing module, S3: integrating and packaging the standardized data in parallel according to time periods through a water conservancy object prediction service module, S4: predicting the water conservancy information at future time points through a water level prediction service module, and S5: after predicting the water conservancy information at future time points several times, the final water conservancy information at the future time points is obtained through iterative calculation by the water level prediction service module.
[0006] Furthermore, the water level prediction service module in S4 predicts the water conservancy information at future time points by integrating the packaged standardized data, and predicts the next future water conservancy information by predicting several times of water conservancy information before the future time point and the predicted water conservancy information, until the time reaches the predicted time point, the water conservancy information at the predicted future time point is replaced by the water conservancy information at the predicted future time point to continue predicting the water conservancy information at the next future time point.
[0007] Furthermore, the water level prediction service module performs water level prediction iterations through Bayesian deep learning, and obtains the time series vector of water flow observed by the hydraulic information model:
[0008] , They are respectively the spatiotemporal data of disaster-causing factors in water conservancy information, constant variable vector, spatial measurement value, initial state vector, general parameter vector, and spatial parameter array. The error is the basic residual neural network deep learning model derived based on the dynamic characteristics of water conservancy information data: ,y m (): model output function, f(): basic residual neural network deep learning model, X: input data, θ: model parameters, ε: noise affected by uncontrollable factors, indicating the random fluctuations between actual observations and predictions that the model fails to fully capture, and the Bayesian posterior probability factor considering time lag:
[0009] , KL divergence is used to measure the distance between distributions and achieve approximate variational optimization: , Y: observation data.
[0010] Furthermore, using the parameter θ Perform Gaussian distribution approximation variation, then the weight and bias distribution at the nth layer of the model for the i-th time is:
[0011] ,
[0012] , μ w and μ b : represents the mean of weights and biases, ρ w and ρ b : represents the variance of weights and biases, and N(0, 1) represents the prior distribution.
[0013] Furthermore, a deep Bayesian network probability vertical prediction model of disaster situation at a certain moment and horizontal time series disaster situation model is constructed. Indicates time point t i The distribution of the joint probability of the set of n influencing factor variables is: , the prior distribution of the parameter is , take the exponential form of prior distribution:
[0014] , , W: equation weight, α: prior distribution hyperfunction that determines weight W, Z(W): normalization constant, E(w) represents the full parameters of the network, and the full parameter values of the network are: , using a dynamic Bayesian network for prediction, the relationship between the continuously updated input value and the predicted value:
[0015] , : Observation results of water level and disaster level at time point N+1, x N+1 : Input value at time point N+1, f(x N+1 , w): prediction function of the model, β: transformation parameter set, S(w): transformation function used to adjust the distribution of weight w, S(w) = αE(w), after determining the predicted model structure and node parameters, use the conditional probability formula Predict the Bayesian probability distribution of water levels and disaster levels at future time points.
[0016] Furthermore, the Bayesian deep learning water level prediction method includes a multi-dimensional multi-step module for storing real-time data and historical data of the water conservancy object information model, and the multi-dimensional multi-step module includes a multi-dimensional cyclic time series and an input of a time series of factors causing disasters: , K: the interval length used when collecting historical data, n: the moment the model is predicting, the water level at which the Bayesian deep learning model is learning predictions: ,y 预测 : predicted water level, y 实测 : Real-time water level, using the previous Second X 输入 Input information, including the previous measured y 实测 Water level information and forecasts 预测 Water level information to predict the next y i+1 Water level information, B i : Forward recursive input information, including historical water level data from jb to j, b: the number of steps to trace back.
[0017] Furthermore, the Bayesian deep learning water level prediction method includes a data table, the water conservancy big data processing module stores the data table, the data table includes a real-time attribute set data table and a historical attribute set data table, the Bayesian deep learning water level prediction method includes a water conservancy prediction client module, and the water conservancy prediction client module is used to access the water conservancy object prediction service module.
[0018] Furthermore, the water level prediction service module includes continuous transmission of data, correction of prediction values, a water level prediction model set based on Bayesian deep learning, and water level prediction effect display. The water level prediction effect display includes a water level prediction effect display graph. The water level prediction service module uses Bayesian deep learning for big data artificial intelligence.
[0019] Furthermore, the Bayesian deep learning water level prediction method includes a water conservancy multi-dimensional multi-step module, and the water conservancy multi-dimensional multi-step module uses a deep learning recurrent neural network RNN and a Bayesian sequence model for repeated training.
[0020] The present invention discloses a multi-dimensional collaborative iterative Bayesian deep learning water level prediction system, which includes a water conservancy big data acquisition module for acquiring water conservancy information, a data storage module for storing water conservancy information, a water conservancy big data processing module for processing water conservancy information into standardized data, a water conservancy object prediction service module for integrating and packaging standardized data according to time points, and a water level prediction service module for predicting water conservancy information at future time points using a Bayesian deep learning algorithm. The water conservancy big data acquisition module is connected to a water conservancy big data processing module, the water conservancy big data processing module is connected to a water conservancy object prediction service module, the water conservancy object prediction service module is connected to a water level prediction service module, and the water conservancy big data processing module and the water conservancy big data acquisition module are connected to a data storage module.
[0021] Beneficial effects of the present invention:
[0022] The multi-dimensional collaborative iterative Bayesian deep learning water level prediction method of the present invention predicts the next future water conservancy information by using the water level prediction service module to predict the water conservancy information of several times before the predicted future time point and the predicted water conservancy information, until the time reaches the predicted time point, the water conservancy information of the predicted future time point is replaced by the water conservancy information of the predicted future time point to iteratively predict the water conservancy information of the next future time point, and then the water conservancy information of the predicted future time point is iteratively calculated by the water level prediction service module after several times of predicting the water conservancy information of the future time point to obtain the final water conservancy information of the future time point. The Bayesian sequence model allows the uncertainty of the model parameters to be quantified, which is achieved through the posterior probability distribution. The Bayesian method provides confidence information about the prediction, making the prediction result more robust. The Bayesian method allows the model to be continuously updated and adjusted as new data is added, thereby improving the prediction performance, and can simultaneously improve the reliability and accuracy of water level prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 A flow chart of a multi-dimensional collaborative iterative Bayesian deep learning water level prediction method in an implementation manner of the present application.
[0024] Figure 2 This is a module schematic diagram of the water level prediction service module in the implementation manner of the present application.
[0025] Figure 3 This is an overall structural block diagram of a multi-dimensional collaborative iterative Bayesian deep learning water level prediction method in an implementation manner of the present application.
[0026] Figure 4 This is a core iterative structural block diagram of a multi-dimensional collaborative iterative Bayesian deep learning water level prediction method in an implementation manner of the present application.
[0027] Figure 5 This is a structural block diagram of a Bayesian deep learning algorithm for a multi-dimensional collaborative iterative Bayesian deep learning water level prediction method in an implementation manner of the present application. DETAILED DESCRIPTION
[0028] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the specific implementation modes of the present invention will be clearly and completely described below.
[0029] The present invention discloses a multi-dimensional collaborative iterative Bayesian deep learning water level prediction method, which includes a water conservancy big data acquisition module, a data storage module, a water conservancy big data processing module, a water conservancy object prediction service module and a water level prediction service module. The water conservancy big data acquisition module is used to acquire water conservancy information, and the water conservancy information includes hydrological, meteorological, geographical and social longitudinal package data. The hydrological data includes rainfall and water level, etc. The meteorological data includes meteorology, rainfall forecast rainfall, rainfall probability, wind direction, temperature, humidity and wind speed, etc., including multiple similar forecast data sources, geographic data includes elevation, geography, evaporation, soil moisture, geological disaster level and terrain assessment, etc., and social longitudinal package data includes building distribution, human activities, public opinion, etc. Water conservancy information realizes the alignment of hydrological, meteorological, geographical and social longitudinal package data through data alignment at the same time, and the user inputs the water conservancy information several hours before the prediction time into the water conservancy big data acquisition module. The data storage module is used to store the water conservancy information acquired by the water conservancy big data acquisition module, and the water conservancy information is stored to facilitate recording and viewing the water conservancy information. The water conservancy big data processing module is used to process the water conservancy information obtained by the water conservancy big data acquisition module into standardized data, and provide standardized data to the data storage module. Processing water conservancy information into standardized data can eliminate data inconsistency, errors and redundancy, ensure data accuracy and integrity, and thus improve data quality. The standardized data format enables different modules to share and exchange water conservancy data more conveniently, and promotes collaboration between modules. The water conservancy object prediction service module is used to receive the access service data provided by the water conservancy big data processing module, and integrate and package the access service data in parallel according to time points. The packaged data can be transmitted at one time, reducing the network overhead and delay caused by multiple transmissions, and improving the speed and efficiency of data transmission. The packaged data as a whole can be more conveniently received and processed by the subsequent water level prediction service module, reducing the tedious operations and data matching work in the data processing process. The water level prediction service module is used to receive the access service data integrated and packaged by the water conservancy object prediction service module to predict water conservancy information at future time points. The water conservancy big data acquisition module is connected to the water conservancy big data processing module, so that the water conservancy big data processing module can process the water conservancy information acquired by the water conservancy big data acquisition module into standardized data. The water conservancy big data processing module is connected to the data storage module, so that the data storage module can receive and store the standardized data provided by the water conservancy big data processing module. The water conservancy big data acquisition module is connected to the data storage module, so that the water conservancy information several hours before the prediction time can be stored. Figure 4As shown, the water conservancy big data processing module is constructed with water conservancy time series data. The water conservancy big data processing module can digitize the water conservancy information transmitted by the big data acquisition module, complete the data, unify the time frequency of each information at hourly resolution, revise, normalize, standardize and segment the data. The advantages of various processing of water conservancy information are: data digitization converts various forms of water conservancy information into numerical values, which is convenient for subsequent mathematical calculations and analysis, and improves the efficiency and accuracy of data processing; data completion makes up for missing values in the data, ensures the integrity of the data, makes subsequent analysis and prediction more reliable, and reduces errors caused by missing data; unified time and frequency are unified at hourly resolution, which helps to compare and analyze different water conservancy information on the same time scale, improves data consistency and comparability; data revision corrects errors and deviations in the data, improves data quality, and lays the foundation for accurate analysis and prediction; normalization and standardization eliminate the influence of data dimension and magnitude, making water conservancy data with different characteristics comparable, facilitating model training and application; data segmentation can divide complex water conservancy data into different stages or categories, which is convenient for targeted analysis and processing, and helps to discover stage laws and characteristics in the data.
[0030] As an implementation method, a multi-dimensional collaborative iterative Bayesian deep learning water level prediction method includes: S1: obtaining water conservancy information data several hours before the prediction time through a water conservancy big data acquisition module, providing sufficient data basis for subsequent analysis and prediction, ensuring the reliability of the prediction, and storing the water conservancy information data for subsequent query, verification and comparative analysis. S2: processing the water conservancy information data into standardized data in a water conservancy time series through a water conservancy big data processing module, and storing the standardized data, eliminating data inconsistency and errors, improving data quality, and thus improving the accuracy of the prediction. The standardized data format facilitates data sharing and exchange between different modules, promotes collaborative work, and stores standardized data for subsequent direct use, thereby improving data processing efficiency. S3: The standardized data is parallelized and integrated and packaged according to time periods through a water conservancy object prediction service module, which facilitates the overall processing of the data by subsequent modules and reduces the complexity of data processing. S4: The water level prediction service module predicts water conservancy information at future time points based on the integrated and packaged standardized data, and predicts based on high-quality and unified format data to improve the credibility of the prediction results. S5: After predicting the water conservancy information at a future time point, the water conservancy information before the predicted future time point and the predicted water conservancy information are used to predict the next future water conservancy information, until the time reaches the predicted time point, the water conservancy information at the predicted future time point is replaced with the water conservancy information at the predicted future time point to iteratively predict the water conservancy information at the next future time point. The prediction results are optimized through continuous iteration, the accuracy and reliability of the prediction are improved, the prediction is updated in real time, so that it can better adapt to the dynamic changes of the water conservancy system, enhance the stability and adaptability of the prediction method, and cope with various complex water conservancy scenarios. After several iterative calculations of the water conservancy information at the predicted future time point, the prediction performance is optimized, and the performance and accuracy of the prediction model can be significantly improved, making it more in line with the actual situation. Data standardization in S2 includes digitization, data completion, unified time and frequency of each information at hourly resolution, revision, normalization, standardization and data segmentation. Numericalization enables data to be quantitatively analyzed and mathematically calculated, which is convenient for subsequent processing using various algorithms and models. Data completion can ensure the integrity of the data, avoid analysis bias and errors caused by missing values, and improve the accuracy and reliability of the prediction. The information is unified at hourly resolution, which facilitates the comparison and comprehensive analysis of water conservancy information from different sources and types on the same time scale, and is conducive to discovering the time patterns and trends in the data. Revision of data can correct errors and deviations in the data, improve data quality, and thus improve the accuracy of subsequent analysis and prediction. Normalization can eliminate the differences in dimensions and magnitudes of different features, making each feature comparable in value, and also help the model converge faster, improving the efficiency of model training and prediction.Standardization makes data conform to standard distribution, further enhances data comparability and universality, reduces the impact of outliers on the model, and improves the stability and generalization of the model. Data segmentation can be used to process and analyze data at different stages or characteristics, and also helps to discover special laws and patterns of data under different time periods or conditions.
[0031] As an implementation method, the water level prediction service module in S4 predicts the water conservancy information at a future time point by integrating the packaged standardized data, and predicts the next future water conservancy information by predicting several times of water conservancy information before the future time point and the predicted water conservancy information, until the time reaches the predicted time point, the water conservancy information at the predicted future time point is replaced by the water conservancy information at the predicted future time point to continue predicting the water conservancy information at the next future time point.
[0032] As an implementation method, the water level prediction service module performs water level prediction iterations through Bayesian deep learning, and the time series vector of water flow observed by the hydraulic information model is:
[0033] The time series vector of water flow refers to the water flow data recorded over a period of time. It can be the change of different types of water flow parameters such as flow, water level, speed, etc. over time. They are respectively the spatiotemporal data of disaster-causing factors in water conservancy information, constant variable vectors, spatial measurements, initial state vectors, general parameter vectors, and spatial parameter arrays. The spatiotemporal data of disaster-causing factors include rainfall, evaporation, groundwater level, topography and other factors. The constant variable vectors represent fixed parameters, such as gravity acceleration and earth rotation rate. The spatial measurements include geographical features such as river width and riverbed slope. The initial state vector represents the water flow state at the starting time. The general parameter vector includes the coefficients and weights of the model. The spatial parameter array represents the spatial change trend or other factors affecting the water flow. The basic residual neural network deep learning model is derived based on the dynamic characteristics of water conservancy information data. ,The basic residual neural network deep learning model is used to process the dynamic characteristics of water conservancy information data,y m () represents the model output function, f() represents the basic residual neural network deep learning model, X represents the input data, θ represents the model parameters, ε represents the noise affected by uncontrollable factors, and the Bayesian posterior probability factor considering the time lag is , KL divergence is used to measure the distance between distributions and achieve approximate variational optimization. , Y represents the observed data, X, Y and parameter θ can be adjusted dynamically, p(θ|y0) is the posterior probability, that is, the probability distribution of parameter θ given the observation result y0, and p(y0|θ) represents the probability of observing y0 given the parameter θ. By minimizing the KL divergence to optimize the model parameters, the true posterior distribution can be effectively approximated to obtain more accurate parameter estimates. KL divergence is used to find the parameter estimate closest to the true distribution. By minimizing the KL divergence, in Bayesian water level prediction, KL divergence is used to find the best model parameters to minimize the prediction error. By comparing the KL divergence under different model structures, the optimal model structure can be selected to improve the prediction accuracy. By dynamically adjusting X, Y and parameter θ, the prediction effect of big data artificial intelligence can be improved. X and Y can be dynamically adjusted through the new multi-source and multi-dimensional uncertainty prediction factor effective information extraction technology, and the parameter θ can be dynamically adjusted through the new deep learning prediction model construction technology. It is expected to greatly improve the prediction performance. Bayesian posterior probability It is used to calculate the probability distribution of parameter θ given the observation result y0. When the observed water flow time series data y0 is available, the belief in the model parameter θ can be updated through Bayes' theorem, which allows the model to adapt to new data more flexibly while retaining the influence of previous knowledge. In Bayesian water level prediction, KL divergence is used to measure the difference between the posterior probability distribution p(θ|y0) and the prior probability distribution p(θ) of the model parameter θ. By minimizing the KL divergence, an approximate posterior probability distribution q(θ) can be found, which is as close to the true posterior probability distribution as possible. The purpose of this is to find an approximate distribution that is easy to handle and close enough to the true distribution when the posterior probability distribution cannot be directly calculated, thereby optimizing the model parameters.
[0034] As an implementation method, cell is an implicit multi-layer DL unit, P is the actual output, c and h are the input and forget gate parameters, and L is the number of units, such as Figure 5 As shown, a deep learning framework with multiple optional technologies such as RNN / LSTM / GRU is used to train the composite model and continuously correct it, so as to build a deterministic deep learning model solution for each prediction point. The Bayesian parameter evaluation method is used to build an uncertainty probability deep learning prediction supplementary model based on historical prior information and parameter terms. Finally, a prediction model library is built to predict the water level and disaster level of multiple small watersheds within 0 to 24 hours and 1 to 7 days.
[0035] As an implementation method, the parameter θ is used to Perform Gaussian distribution approximation variation, then the weight and bias distribution at the nth layer of the model for the i-th time is:
[0036] ,
[0037] , μ w and μ b represents the mean of weights and biases, ρ w and ρ b represents the variance of weights and biases, and N(0, 1) represents the prior distribution.
[0038] As an implementation method, a deep Bayesian network probability vertical prediction model of disaster situation at a certain moment and horizontal time series disaster situation model is constructed. Indicates time point t i The distribution of the joint probability of the set of n influencing factor variables is: , the prior distribution of the parameter is , take the exponential form of prior distribution:
[0039] , , W represents the equation weight, α represents the prior distribution hyperfunction that determines the weight W, Z(W) represents the normalization constant, ensuring that the sum of the probability density function is equal to 1, E(w) represents the full parameters of the network, and the full parameter values of the network are: , using a dynamic Bayesian network for prediction, the relationship between the continuously updated input value and the predicted value: , represents the observation results of water level and disaster level at time point N+1, x N+1 represents the input value at time point N+1, f(x N+1 , w) represents the prediction function of the model, which is used to predict the result at the next time point, β represents the transformation parameter set, which is used to adjust the shape of S(w), S(w) represents the transformation function used to adjust the distribution of weight w, S(w) = αE(w), after determining the predicted model structure and node parameters, use the conditional probability formula Predict the Bayesian probability distribution of water level and disaster level at future time points. Use the determined model structure and node parameters to make predictions. For water level prediction, use historical observation data and current input values to predict future water levels, use dynamic Bayesian networks to simulate the relationship between input values and predicted values, and continuously update model parameters to adapt to new observation data. By calculating f(x N+1 ,w)exp(-S(w))dw, we can get the prediction result at the next time point, that is, we can use the prior knowledge and observation data to update the posterior distribution of the model parameters to obtain a more accurate and reliable prediction model, which not only considers the influence of historical data but also includes prior knowledge. After determining the prediction model structure and node parameters, we use the conditional probability formula The Bayesian probability distribution of water level and epidemic level at future time points is used to predict the conditional probability formula. Given the current water level and epidemic level and related factors F t Under the condition of flooding, the probability distribution of water level and epidemic level at the next time point is predicted, so as to obtain a continuous time series prediction model, which can take into account the influence of various factors and make reasonable predictions for future situations.
[0040] As an implementation method, the Bayesian deep learning water level prediction method includes a multi-dimensional multi-step module for storing real-time data and historical data of a water conservancy object information model, and the multi-dimensional multi-step module includes a multi-dimensional cyclic time series and an input of a time series of factors causing disasters:
[0041] , K represents the interval length used when collecting historical data, n: the water level of the Bayesian deep learning model for learning prediction at the moment the model is predicting:
[0042] ,y 预测 : predicted water level, y 实测 : Real-time water level, using the previous Second X 输入 Input information, including the previous measured y 实测 Water level information and forecasts 预测 Water level information to predict the next y i+1 Water level information, B i : forward recursive input information, including historical water level data from jb to j, b: the number of steps back. 输入 and 实测 Information about 预测 The next water level information y i+1 water level for training and prediction, and can apply real-time i+1 The prediction accuracy and prediction error of the prediction model are calibrated by the measured water level. In this way, in the continuous training prediction, only the water level information of the last prediction is not measured, but the other information is real-time and accurate, which effectively improves the reliability and real-time performance of the input data. In addition, in the continuous multi-step prediction calculation, the reliability of the input data is further improved by continuously iterating and using the real-time predicted water level information to replace the water level information on the next side of the prediction.
[0043] As an implementation method, the water level prediction service module includes continuous data transmission, correction of prediction values, a water level prediction model set based on Bayesian deep learning, and water level prediction effect display. The water level prediction effect display includes a water level prediction effect display diagram. The water level prediction service module is used for big data artificial intelligence through Bayesian deep learning. Continuous data transmission can ensure the consistency of data, so that water level prediction can be based on continuous data streams, avoiding prediction deviations caused by data interruptions. Correction of prediction values can improve prediction accuracy, reduce prediction errors, and make prediction results closer to actual water levels. The water level prediction model set based on Bayesian deep learning can improve prediction accuracy, use the powerful ability of deep learning to mine complex patterns and relationships in data, and can process various types and characteristics of water conservancy data to adapt to complex water conservancy environments. The water level prediction effect display diagram can present the prediction results in an intuitive way, which is easy to understand and analyze, and is convenient for evaluating and comparing different prediction models or methods, which helps to optimize the prediction strategy. It is used for big data artificial intelligence to process massive data, make full use of a large amount of water conservancy data, mine potential valuable information, and use the ability of artificial intelligence to achieve automated analysis and prediction, reducing manual intervention and errors.
[0044] As an implementation method, the Bayesian deep learning water level prediction method includes a data table. The water conservancy big data processing module stores the data table. The data table includes a real-time attribute set data table and a historical attribute set data table. The Bayesian deep learning water level prediction method includes a water conservancy prediction client module. The water conservancy prediction client module is used to access the water conservancy object prediction service module.
[0045] As an implementation method, the Bayesian deep learning water level prediction method includes a water conservancy multi-dimensional multi-step module. The water conservancy multi-dimensional multi-step module stores the real-time data processed by the water conservancy big data processing module and stores the data in a data table. The water level prediction service module reads the water conservancy multi-dimensional multi-step data and continuously outputs the water level data by hour. The water conservancy multi-dimensional multi-step module uses a deep learning recurrent neural network RNN and a Bayesian sequence model for repeated training and verification. The recurrent neural network RNN is able to process sequence data, which allows the model to use past data to predict future water level changes, which is very suitable for the characteristics of hydrological sequence data. RNN can capture long-term dependencies, which is crucial for predicting water level changes over a long span. The Bayesian sequence model allows the uncertainty of the model parameters to be quantified through the posterior probability distribution. The Bayesian method provides confidence information about the prediction, making the prediction results more robust. The Bayesian method allows the model to be continuously updated and adjusted as new data is added, thereby improving the prediction performance.
[0046] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the protection scope of the appended claims of the present invention.
Claims
1. A multi-dimensional collaborative iterative Bayesian deep learning water level prediction method, It is characterized in that include: S1: Acquire water conservancy information data several hours before the prediction time through the water conservancy big data acquisition module; S2: Process the water conservancy information data into standardized data in water conservancy time series through the water conservancy big data processing module; S3: Through the water conservancy object prediction service module, the standardized data is parallelized and integrated and packaged according to time periods; S4: The water level prediction service module predicts water conservancy information at future time points; S5: After predicting the water conservancy information at future time points for several times, the water level prediction service module iteratively calculates and obtains the final water conservancy information at the future time point; The water level prediction service module uses Bayesian deep learning to iterate water level predictions, and the time series vector of water flow observed by the water conservancy information model is: ； are the spatiotemporal data of disaster factors in water conservancy information, constant variable vector, spatial measurement value, initial state vector, general parameter vector, and spatial parameter array, respectively. ε is the error; Based on the dynamic characteristics of water conservancy information data, the basic residual neural network deep learning model is derived: ； y m (): model output function; f(): basic residual neural network deep learning model; X: input data; θ: model parameters; ε: noise caused by uncontrollable factors; Bayesian posterior probability factor considering time lag: ； p(θ|y 0 ): represents the posterior probability; p(θ): represents the prior probability distribution of model parameter θ; p(y 0 |θ): indicates that y is observed under the condition of given parameter θ 0 The probability of KL divergence is used to measure the distance between distributions and achieve approximate variational optimization: ； Y: Observation data; Using the parameter θ Perform Gaussian distribution approximation variation, then the weight and bias distribution at the nth layer of the model for the i-th time is: ； ； μ w and μ b : represents the mean of weights and biases; ρ w and ρ b : represents the variance of weights and biases; N (0, 1): represents the prior distribution; q(θ): represents the approximate posterior probability distribution; Construct a deep Bayesian network probability vertical prediction moment disaster situation and horizontal time series disaster situation model. Indicates time point t i The distribution of the joint probability of the set of n influencing factor variables is: ； The prior distribution of the parameter is , take the exponential form of prior distribution: ； ； W: equation weight; α: The hyperfunction of the prior distribution that determines the weight W; Z(W): normalization constant; E(w) represents the full parameters of the network, and the full parameter values of the network are: ； Use dynamic Bayesian networks to make predictions, and continuously update the relationship between input values and predicted values: ； : Observation results of water level and disaster level at time point N+1; x N+1 : Input value at time point N+1; f(x N+1 ,w): prediction function of the model; β: transformation parameter set; S(w): Transformation function used to adjust the distribution of weight w, S(w) = αE(w); After determining the predicted model structure and node parameters, use the conditional probability formula Predict the Bayesian probability distribution of water level and disaster severity at future time points; F t : Related factors.
2. According to the multi-dimensional collaborative iterative Bayesian deep learning water level prediction method described in claim 1, Features: The water level prediction service module in S4 predicts the water conservancy information at future time points by integrating the packaged standardized data, and predicts the next future water conservancy information by predicting several times of water conservancy information before the future time point and the predicted water conservancy information, until the time reaches the predicted time point, the water conservancy information at the predicted future time point is replaced by the water conservancy information at the predicted future time point to continue predicting the water conservancy information at the next future time point.
3. According to the multi-dimensional collaborative iterative Bayesian deep learning water level prediction method described in claim 1, Features: The Bayesian deep learning water level prediction method includes a multi-dimensional multi-step module for storing real-time data and historical data of a water conservancy object information model, and the multi-dimensional multi-step module includes a multi-dimensional cyclic time series; Input of time series of factors causing disasters: ； K: the interval length used when collecting historical data; n: the moment the model is predicting; The water level of Bayesian deep learning model for learning prediction: ； y 预测 : Predict water level; y 实测 : Real-time water level; Before adoption Second X 输入 Input information, including the previous measured y 实测 Water level information and forecasts 预测 Water level information to predict the next y i+1 Water level information; B i : Forward recursive input information, including historical water level data from jb to j; b: The number of steps backtracked.
4. According to the multi-dimensional collaborative iterative Bayesian deep learning water level prediction method described in claim 1, Features: The Bayesian deep learning water level prediction method includes a data table, and the water conservancy big data processing module stores the data table, and the data table includes a real-time attribute set data table and a historical attribute set data table; The Bayesian deep learning water level prediction method includes a water conservancy prediction client module, which is used to access the water conservancy object prediction service module.
5. According to the multi-dimensional collaborative iterative Bayesian deep learning water level prediction method described in claim 1, Features: The water level prediction service module includes continuous data transmission, correction of prediction values, a water level prediction model set based on Bayesian deep learning, and water level prediction effect display, which includes a water level prediction effect display diagram; The water level prediction service module is used for big data artificial intelligence through Bayesian deep learning.
6. According to the multi-dimensional collaborative iterative Bayesian deep learning water level prediction method described in claim 1, Features: The Bayesian deep learning water level prediction method includes a water conservancy multi-dimensional multi-step module, which uses a deep learning recurrent neural network RNN and a Bayesian sequence model for repeated training.
7. A multi-dimensional collaborative iterative Bayesian deep learning water level prediction system, Features A Bayesian deep learning water level prediction method using any one of multi-dimensional collaborative iterations as claimed in claims 1-6, comprising: Water conservancy big data acquisition module: used to obtain water conservancy information data several hours before the prediction time; Water conservancy big data processing module: used to process water conservancy information data into standardized data in water conservancy time series; Water conservancy object prediction service module: used to integrate and package standardized data in parallel according to time periods; Water level prediction service module: used to predict water conservancy information at future time points; The water conservancy big data acquisition module establishes a connection with the water conservancy big data processing module, the water conservancy big data processing module establishes a connection with the water conservancy object prediction service module, and the water conservancy object prediction service module establishes a connection with the water level prediction service module.