Runoff prediction method based on multi-source data fusion

Abstract

The invention discloses a runoff prediction method based on multi-source data fusion, and relates to the technical field of model prediction, and the method comprises the following steps: S1, data source selection and collection; s2, data preprocessing; s3, designing a data fusion strategy; s4, feature extraction and selection; s5, model selection and construction; s6, model training and verification; s7, optimizing and improving the model; and S8, runoff prediction and result analysis. According to the runoff prediction method based on multi-source data fusion, through fusion of multi-source data and introduction of multiple deep learning models, the capture ability of the model to a complex non-linear relationship is enhanced, and the physical consistency and the coping ability to an extreme event are also ensured; a more comprehensive runoff prediction result is provided, the prediction robustness and the reliability of an analysis result can be improved, and the defects that a traditional model is single in data source, poor in physical consistency, insufficient in prediction uncertainty analysis and the like are overcome.

Description

Technical Field

[0001] The invention relates to the technical field of model prediction, and in particular to a runoff prediction method based on multi-source data fusion. Background Art

[0002] Runoff prediction is of great significance in many fields such as water resources management, flood prevention and disaster reduction, and agricultural irrigation. Traditional runoff prediction methods mainly rely on physical models and statistical models, which usually use a single data source, such as precipitation, topographic data, and hydrological station observation data. However, due to the complexity of the natural environment and the high uncertainty of meteorological changes, the limitations of a single data source are obvious, and it is impossible to fully capture the multidimensional factors affecting runoff. In addition, with the development of big data technology, the widespread application of multi-source data such as remote sensing data, meteorological forecast data, and geographic information system (GIS) data has brought new opportunities for runoff prediction. These multi-source data can make up for the shortcomings of traditional data sources and improve the accuracy and robustness of runoff prediction through data fusion.

[0003] Among the existing multi-source data fusion runoff prediction methods, the combination of physical models and data-driven models is one of the mainstream trends. However, traditional model building methods have shown certain limitations when dealing with complex nonlinear relationships and high-dimensional data. Especially in extreme weather events and data-scarce areas, the generalization ability and adaptability of the model are poor, which can easily lead to deviations in prediction results. In recent years, with the development of machine learning technology, data-driven models (such as deep learning) have gradually been applied to runoff prediction, but these models usually rely only on large amounts of data for training and lack an understanding of physical processes, which may lead to unreasonable prediction results.

[0004] There are several main disadvantages in the prior art:

[0005] Single data source: Many traditional methods rely on a single data source, which makes it impossible to fully utilize multidimensional data to improve the accuracy and stability of predictions.

[0006] Poor physical consistency of the model: Although data-driven models perform well in fitting ability, due to the lack of understanding of physical processes, the predicted results may sometimes not conform to physical laws.

[0007] Inadequate ability to handle extreme weather events: Existing models perform poorly when dealing with extreme weather events and are prone to large prediction errors.

[0008] Excessive reliance on key variables: Traditional methods tend to rely heavily on certain key variables and are easily affected by noise and errors in the input variables, resulting in unstable predictions. Summary of the invention

[0009] In view of the deficiencies in the prior art, the present invention provides a runoff prediction method based on multi-source data fusion to solve the problems raised in the above background technology.

[0010] To achieve the above object, the present invention provides the following technical solution: a runoff prediction method based on multi-source data fusion, comprising the following steps:

[0011] S1. Selection and collection of data sources

[0012] Select multiple traditional data sources suitable for runoff prediction, including meteorological data, hydrological data, geographic data, and remote sensing data. At the same time, introduce non-traditional data sources, including social media data and crowd science data, to form multi-source data; and use drones for real-time observation to provide high-resolution ground information to enhance the understanding of terrain and vegetation cover;

[0013] S2. Data preprocessing

[0014] Preprocess multi-source data, including removing outliers, filling missing values, and standardizing data. For remote sensing data, image processing techniques such as cloud removal and noise filtering are used to improve data reliability due to factors such as cloud cover and sensor failure.

[0015] S3. Design of data fusion strategy

[0016] Design strategies suitable for multi-source data fusion, adopt a multi-level multi-source data fusion framework, and make comprehensive use of data from different sources based on the semantic level of the data during fusion;

[0017] S4. Feature extraction and selection

[0018] Extract relevant features from the fused data, including precipitation intensity, temperature change rate, surface vegetation index, and watershed area; and select and reduce the dimension of the extracted features to avoid feature redundancy and overfitting;

[0019] S5. Model selection and construction

[0020] Select appropriate prediction models according to the needs of runoff prediction, build a deep learning model of multi-model combination based on physical constraints, and introduce mixed reality technology to build a "virtual watershed". Combine simulation experiments with actual data to improve the prediction ability of the model;

[0021] S6. Model training and validation

[0022] The preprocessed data is input into the model for training, and two training models, a learning model and a reinforcement model, are used to evaluate the performance of the model. An adversarial training strategy is used to enhance the model's ability to predict extreme weather events when generating adversarial samples.

[0023] S7. Model optimization and improvement

[0024] After the initial training and validation of the model, the model is further optimized;

[0025] S8. Runoff prediction and result analysis

[0026] The optimized model is used to predict runoff in future time periods, and the accuracy and uncertainty of the prediction results are analyzed. The prediction results are combined with actual hydrological and meteorological conditions for reasonable interpretation and application to support water resources management and disaster prevention and mitigation work.

[0027] To further optimize the technical solution, in step S1:

[0028] Among traditional data sources, meteorological data include precipitation, temperature, and humidity; hydrological data include river flow and water level; geographic data include topography and land use type; remote sensing data include satellite images and vegetation coverage;

[0029] Among non-traditional data sources, social media data and crowd science data, real-time precipitation or water level change information in a specific area can be obtained by analyzing weather-related discussions, pictures and videos on social media; this can be used as supplementary information to traditional data sources, especially in areas where traditional data sources are lacking.

[0030] To further optimize the technical solution, in step S2, the preprocessing of multi-source data is based on an adaptive data cleaning algorithm, which automatically selects the best data processing method according to different data sources and data types, including:

[0031] Use autoencoder networks in deep learning to automatically detect and repair abnormal data;

[0032] Generate reasonable filling of missing data through generative adversarial networks;

[0033] A data source traceability system is established based on blockchain technology to verify the authenticity and reliability of the data.

[0034] To further optimize the technical solution, in step S3, when the semantic level of data is considered during fusion, the following process is specifically included:

[0035] First, low-level physical fusion of meteorological and hydrological data and medium-level physical fusion of social media and crowd science data are performed;

[0036] Then, high-level feature fusion is performed in combination with geographic information data;

[0037] Second, use graph neural networks to process complex geospatial relationships, or use self-attention mechanisms to capture long-distance dependencies between different data sources;

[0038] Finally, by introducing federated learning technology, data privacy can be protected while data is shared between different organizations, further enriching the diversity of data sources.

[0039] To further optimize the technical solution, in step S4, transfer learning and meta-learning techniques are also introduced to transfer learning feature extractors from similar hydrological prediction tasks, or use a meta-learning framework to quickly adjust the model to adapt to new data distribution;

[0040] Image processing technology is used to extract complex spatial features from remote sensing data, or natural language processing technology is used to extract weather-related emotions and keywords from social media texts as auxiliary features.

[0041] To further optimize the technical solution, in step S5, the construction of the deep learning model of the multi-model combination includes the following specific processes:

[0042] Define the physical equations and select the physical equations suitable for the runoff process, including Saint-Venant equations, Navier-Stokes equations or rainfall-runoff model;

[0043] Deep learning model design, building a deep neural network whose input is the features of multi-source data and output is runoff;

[0044] Definition of loss function,The loss function consists of two parts: one is the traditional supervised learning loss, which is used to minimize the error between the model prediction value and the true observation value; the other is the physical constraint loss, which is used to ensure that the model prediction value satisfies the physical equation; the physical constraint loss is represented by the residual of the partial differential equation;

[0045] Model training, using gradient descent method to train neural network parameters and correct the prediction results through physical equations;

[0046] Model validation and tuning: evaluate model performance on the validation set to ensure that the model can not only accurately predict runoff but also meet the constraints of physical laws. Based on the validation results, adjust the network structure or optimization method to improve the model's predictive performance.

[0047] To further optimize the technical solution, in step S6, the learning model and the reinforcement model respectively include the following contents:

[0048] In the learning model, assume an input data set (X) that has no labels, design a learning task, and learn an effective representation (Z) of the input data by constructing a proxy task;

[0049] Choose a proxy task, such as time series forecasting, spatiotemporal segment sorting, or masked data recovery tasks;

[0050] For the time series data in runoff prediction (X = {x1, x2, ..., x T}), design a time segment prediction task:

[0051]

[0052] in, is the predicted data at the (t+k)th moment, (f θ ) is a deep learning model with parameters (θ);

[0053] Use model (f θ ) is trained by minimizing the loss function of the proxy task (L SSL ), and learn an effective representation (Z) of the input data:

[0054] Z=f θ (X)

[0055] The loss function is defined as:

[0056]

[0057] Where (N) is the number of training samples, is the predicted value of the (i)th sample, is the corresponding true value;

[0058] In the reinforcement model, in the framework of reinforcement learning, an environment (E) and an agent (A) are defined. The agent learns the optimal strategy (π(a|s)) by interacting with the environment, where (s) is the state and (a) is the action;

[0059] The key variables in the runoff prediction task, such as precipitation, evaporation, and soil moisture, are taken as states (s t ):

[0060] s t =[Precipitation t , evaporation t , soil temperature t , ...]

[0061] The action chosen by the agent (a t ) is the adjustment of model parameters, feature selection or change of prediction strategy, adjusting the learning rate of the model or choosing to use a specific feature:

[0062] a t=adjust(learning rate)or select(feature subset)

[0063] Design a reward function (r t ), so that the agent chooses actions that maximize the accuracy of runoff prediction, for example, based on the negative of the prediction error as a reward:

[0064]

[0065] Among them, (y t ) is the actual runoff, is the predicted value of the model;

[0066] Use the policy gradient method to optimize the strategy (π(a|s)) so that the cumulative reward maximize:

[0067]

[0068] To further optimize the technical solution, in step S7, the optimized contents include:

[0069] Reselect or combine features and adjust the model structure, such as adding hidden layers and changing activation functions;

[0070] Introducing new data sources, such as real-time remote sensing data, to improve the prediction accuracy of the model;

[0071] An adaptive optimization algorithm is used to automatically adjust model parameters to improve the adaptability and predictive ability of the model, and the quantum optimization algorithm in quantum computing is used to accelerate the optimization process of the model and obtain better prediction results in a shorter time.

[0072] Further optimizing the technical solution, in step S8, in the runoff prediction, multiple possible runoff prediction scenarios are generated based on the deep learning model, and the confidence interval of each scenario is calculated;

[0073] By modeling the probability distribution of network weights, multiple possible prediction results are generated. Assume that the output of the runoff prediction model is The input is (X t ), including precipitation, temperature, and terrain characteristics, the deep learning model is expressed as:

[0074]

[0075] Where (θ) is the weight of the neural network, is the training data set;

[0076] Use variational inference or Markov Chain Monte Carlo methods to approximate the posterior distribution The loss function is defined as:

[0077]

[0078] Here, (q(θ)) is an approximation to the posterior distribution and (KL) is the Kullback-Leibler divergence, which measures the difference between two distributions.

[0079] Further optimizing the technical solution, in step S8, in the runoff prediction, the influence of different input variables on the prediction results is determined by a sensitivity analysis model, which helps to identify key environmental factors and potential risks;

[0080] Assume that the output of runoff prediction is The input variables include multiple features It is used to evaluate the impact of each input variable on the output by calculating the partial derivative of the model output with respect to the input:

[0081]

[0082] Among them, (S i ) represents the input variable sensitivity to model output;

[0083] By calculating the sensitivity of each input variable (S i ), rank the importance of variables and identify the most critical factors for runoff prediction for model optimization or result support.

[0084] Compared with the prior art, the present invention provides a runoff prediction method based on multi-source data fusion, which has the following beneficial effects:

[0085] This runoff prediction method based on multi-source data fusion, by fusing multi-source data and introducing a variety of deep learning models, not only enhances the model's ability to capture complex nonlinear relationships, but also ensures physical consistency and the ability to respond to extreme events; it provides more comprehensive runoff prediction results, helps to improve the robustness of the prediction and the reliability of the analysis results, overcomes the shortcomings of traditional models in terms of single data source, poor physical consistency, and insufficient prediction uncertainty analysis, and provides a more accurate and stable prediction tool for water resources management, disaster prevention and mitigation and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 A schematic diagram of a flow chart of a runoff prediction method based on multi-source data fusion proposed by the present invention;

[0087] Figure 2 A data classification diagram of traditional data sources in a runoff prediction method based on multi-source data fusion proposed by the present invention;

[0088] Figure 3 A flowchart of a deep learning model of a multi-model combination in step S5 of a runoff prediction method based on multi-source data fusion proposed by the present invention;

[0089] Figure 4 This is a flow chart of constructing a "virtual watershed" in a runoff prediction method based on multi-source data fusion proposed by the present invention. DETAILED DESCRIPTION

[0090] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0091] Embodiment 1:

[0092] See also Figure 1 , a runoff prediction method based on multi-source data fusion, including the following steps:

[0093] S1. Selection and collection of data sources

[0094] Select multiple traditional data sources suitable for runoff prediction, including meteorological data, hydrological data, geographic data, and remote sensing data. At the same time, introduce non-traditional data sources, including social media data and crowd science data, to form multi-source data. Use drones for real-time observations to provide high-resolution ground information to enhance the understanding of terrain and vegetation cover.

[0095] In this embodiment, if Figure 2 As shown, among the traditional data sources, meteorological data include precipitation, temperature, and humidity; hydrological data include river flow and water level; geographic data include topography and land use type; remote sensing data include satellite images and vegetation coverage;

[0096] Among non-traditional data sources, social media data and crowd science data, by analyzing weather-related discussions, pictures and videos on social media, real-time precipitation or water level change information in a specific area can be obtained; it is used as supplementary information for traditional data sources, especially in areas where traditional data sources are lacking, and has important application value.

[0097] S2. Data preprocessing

[0098] Preprocess multi-source data, including removing outliers, filling missing values, and standardizing data. For remote sensing data, image processing techniques such as cloud removal and noise filtering are used to improve data reliability due to factors such as cloud cover and sensor failure.

[0099] In this embodiment, the preprocessing of multi-source data is based on an adaptive data cleaning algorithm, which automatically selects the best data processing method according to different data sources and data types, including:

[0100] Use autoencoder networks in deep learning to automatically detect and repair abnormal data;

[0101] Generate reasonable filling of missing data through generative adversarial networks;

[0102] A data source traceability system is established based on blockchain technology to verify the authenticity and reliability of the data.

[0103] S3. Design of data fusion strategy

[0104] Design a strategy suitable for multi-source data fusion, adopt a multi-level and multi-source data fusion framework, and comprehensively utilize data from different sources based on the semantic level of the data during fusion.

[0105] In this embodiment, when the semantic level of data is considered during fusion, the following process is specifically included:

[0106] First, low-level physical fusion of meteorological and hydrological data and mid-level information fusion of social media and crowd science data are performed;

[0107] Then, high-level feature fusion is performed in combination with geographic information data;

[0108] Second, use graph neural networks to process complex geospatial relationships, or use self-attention mechanisms to capture long-distance dependencies between different data sources;

[0109] Finally, by introducing federated learning technology, data privacy can be protected while data is shared between different organizations, further enriching the diversity of data sources.

[0110] in:

[0111] Social media data: First, natural language processing (NLP) is performed on social media data, including sentiment analysis, keyword extraction, and text classification, to extract substantive information related to weather or hydrological conditions. In addition, computer vision technology can be used to analyze pictures and videos uploaded by users to identify features such as precipitation conditions and water level changes.

[0112] Crowd science data: These data may come from non-professionals, so data cleaning and verification are needed to ensure the accuracy of the data. Crowdsourcing verification mechanisms can be used to combine multiple user reports to verify observation data in the same area, thereby improving the reliability of the data.

[0113] Drone observation data: Drone data has the advantage of high resolution, but the data volume is large and complex. Image processing algorithms are needed to perform terrain analysis and vegetation cover identification, and compare with traditional remote sensing data to ensure the consistency of observation results.

[0114] Social media data and crowd science data usually have low spatiotemporal accuracy, so spatiotemporal alignment is required. These data can be mapped to the same time scale and spatial range as traditional data sources by constructing time windows and spatial grids. Drone data is often used as a reference due to its high accuracy.

[0115] S4. Feature extraction and selection

[0116] Relevant features are extracted from the fused data, including precipitation intensity, temperature change rate, surface vegetation index, and watershed area. The extracted features are selected and reduced in dimension to avoid feature redundancy and overfitting.

[0117] In this embodiment, transfer learning and meta-learning techniques are also introduced to transfer learning feature extractors from similar hydrological prediction tasks, or use a meta-learning framework to quickly adjust the model to adapt to new data distributions;

[0118] Image processing technology is used to extract complex spatial features from remote sensing data, or natural language processing technology is used to extract weather-related emotions and keywords from social media texts as auxiliary features.

[0119] S5. Model selection and construction

[0120] Select the appropriate prediction model according to the needs of runoff prediction, build a deep learning model with multiple model combinations based on physical constraints, introduce mixed reality technology to build a "virtual watershed", and combine simulation experiments with actual data to improve the prediction ability of the model.

[0121] like Figure 3 As shown, in this embodiment, the construction of the deep learning model of the multi-model combination includes the following specific processes:

[0122] Define the physical equations and select the physical equations suitable for the runoff process, including Saint-Venant equations, Navier-Stokes equations or rainfall-runoff model;

[0123] Deep learning model design, building a deep neural network whose input is the features of multi-source data and output is runoff;

[0124] Definition of loss function,The loss function consists of two parts: one is the traditional supervised learning loss, which is used to minimize the error between the model prediction value and the true observation value; the other is the physical constraint loss, which is used to ensure that the model prediction value satisfies the physical equation; the physical constraint loss is represented by the residual of the partial differential equation;

[0125] Model training, using gradient descent method to train neural network parameters and correct the prediction results through physical equations;

[0126] Model validation and tuning: evaluate model performance on the validation set to ensure that the model can not only accurately predict runoff but also meet the constraints of physical laws. Based on the validation results, adjust the network structure or optimization method to improve the model's predictive performance.

[0127] S6. Model training and validation

[0128] The preprocessed data is input into the model for training, and two training models, a learning model and a reinforcement model, are used to evaluate the performance of the model. An adversarial training strategy is adopted to enhance the model's ability to predict extreme weather events when generating adversarial samples.

[0129] In this embodiment, the learning model and the reinforcement model respectively include the following contents:

[0130] In the learning model, assume an input data set (X) that has no labels, design a learning task, and learn an effective representation (Z) of the input data by constructing a proxy task;

[0131] Choose a proxy task, such as time series forecasting, spatiotemporal segment sorting, or masked data recovery tasks;

[0132] For the time series data in runoff prediction (X = {x1, x2, ..., x T}), design a time segment prediction task:

[0133]

[0134] in, is the predicted data at the (t+k)th moment, (f θ ) is a deep learning model with parameters (θ);

[0135] Use model (f θ ) is trained by minimizing the loss function of the proxy task (L SSL ), and learn an effective representation (Z) of the input data:

[0136] Z=f θ (X) The loss function is defined as:

[0137]

[0138] Where (N) is the number of training samples, is the predicted value of the (i)th sample, is the corresponding true value.

[0139] When the model is used, it is first trained on unlabeled data to generate proxy tasks and optimize representation learning. Subsequently, these representations (Z) will be used as input features for the downstream runoff prediction task. In this way, even in the absence of a large amount of labeled data, the model can obtain useful information in the data through self-supervised learning.

[0140] In the reinforcement model, in the framework of reinforcement learning, an environment (E) and an agent (A) are defined. The agent learns the optimal strategy (π(a|s)) by interacting with the environment, where (s) is the state and (a) is the action;

[0141] The key variables in the runoff prediction task, such as precipitation, evaporation, and soil moisture, are taken as states (s t ):

[0142] s t =[Precipitation t , evaporation t , soil temperature t , ...]

[0143] The action chosen by the agent (a t ) is the adjustment of model parameters, feature selection or change of prediction strategy, adjusting the learning rate of the model or choosing to use a specific feature:

[0144] a t =adjust(learning rate)or select(feature subset)

[0145] Design a reward function (r t ), so that the agent chooses actions that maximize the accuracy of runoff prediction, for example, based on the negative of the prediction error as a reward:

[0146]

[0147] Among them, (y t ) is the actual runoff, is the predicted value of the model;

[0148] Use the policy gradient method to optimize the strategy (π(a|s)) so that the cumulative reward maximize:

[0149]

[0150] In the runoff prediction task, the agent optimizes the prediction strategy by continuously interacting with the environment. The reinforcement model has advantages in dealing with runoff prediction in a dynamic environment, and can gradually improve the prediction ability through trial and error learning. Especially when extreme weather events occur, the agent can adjust the prediction strategy to deal with abnormal situations, thereby improving the robustness of the model.

[0151] In this embodiment, an adversarial training strategy is adopted to enhance the model's ability to predict extreme weather events when generating adversarial samples. Combining learning and reinforcement models can further enhance the model's ability to predict extreme weather events. In particular, by introducing adversarial training, the model can be trained on adversarial samples, thereby enhancing the robustness of the model.

[0152] Generate adversarial examples using generative adversarial networks (GANs) or gradient-based methods that simulate extreme weather events by perturbing the input data.

[0153] Self-supervised learning models and reinforcement learning models are trained on the generated adversarial samples to enhance the adaptability and predictive ability of the models in extreme environments.

[0154] Reinforcement learning agents are trained in adversarial environments. Through adversarial interactions with the environment, agents are better able to cope with uncertainty and extreme situations and optimize the final prediction strategy.

[0155] In actual prediction, using self-supervised learning and reinforcement learning models enhanced by adversarial training, the model can better predict runoff under extreme weather events. Especially in the face of extreme events, the model can adjust its strategy in time and provide more robust prediction results, thus helping to deal with natural disasters and risk management.

[0156] S7. Model optimization and improvement

[0157] After the initial training and verification of the model, the model is further optimized.

[0158] In this embodiment, the optimized contents include:

[0159] Reselect or combine features and adjust the model structure, such as adding hidden layers and changing activation functions;

[0160] Introducing new data sources, such as real-time remote sensing data, to improve the prediction accuracy of the model;

[0161] An adaptive optimization algorithm is used to automatically adjust model parameters to improve the adaptability and predictive ability of the model, and the quantum optimization algorithm in quantum computing is used to accelerate the optimization process of the model and obtain better prediction results in a shorter time.

[0162] S8. Runoff prediction and result analysis

[0163] The optimized model is used to predict runoff in future time periods, and the accuracy and uncertainty of the prediction results are analyzed. The prediction results are combined with actual hydrological and meteorological conditions for reasonable interpretation and application to support water resources management and disaster prevention and mitigation work.

[0164] In this embodiment, in runoff prediction, multiple possible runoff prediction scenarios are generated based on the deep learning model, and the confidence interval of each scenario is calculated;

[0165] By modeling the probability distribution of network weights, multiple possible prediction results are generated. Assume that the output of the runoff prediction model is The input is (X t ), including precipitation, temperature, and terrain characteristics, the deep learning model is expressed as:

[0166]

[0167] Where (θ) is the weight of the neural network, is the training data set;

[0168] Use variational inference or Markov Chain Monte Carlo methods to approximate the posterior distribution The loss function is defined as:

[0169]

[0170] Here, (q(θ)) is an approximation to the posterior distribution and (KL) is the Kullback-Leibler divergence, which measures the difference between two distributions.

[0171] In runoff prediction, the learning model will generate a set of possible runoff prediction values ​​for each time point, which correspond to different sampling results of the model weights. By performing statistical analysis on these prediction values, the confidence interval for each time point is calculated. This method ensures that the prediction results not only contain the most likely runoff value, but also provide a range of uncertainty, providing more reference information for decision makers.

[0172] In this embodiment, in runoff prediction, the sensitivity analysis model is used to determine the impact of different input variables on the prediction results, which helps to identify key environmental factors and potential risks;

[0173] Assume that the output of runoff prediction is The input variables include multiple features It is used to evaluate the impact of each input variable on the output by calculating the partial derivative of the model output with respect to the input:

[0174]

[0175] Among them, (S i ) represents the input variable sensitivity to model output;

[0176] By calculating the sensitivity of each input variable (S i ), rank the importance of variables and identify the most critical factors for runoff prediction for model optimization or result support.

[0177] Sensitivity analysis helps identify which input variables (such as precipitation, evaporation, soil moisture, etc.) have a significant impact on the model output in runoff prediction. By analyzing the sensitivity of these variables, the model can be optimized for key factors and focused on monitoring these variables in practical applications to improve the accuracy and reliability of runoff prediction.

[0178] Embodiment 2:

[0179] Based on the runoff prediction method based on multi-source data fusion described in Example 1, Figure 4 As shown in the figure, constructing a "virtual watershed" includes the following processes:

[0180] Data Collection and Processing

[0181] Collect multi-source data in the watershed, including topography, soil, vegetation, meteorology and hydrology data. These data will serve as the basis for the construction of the virtual watershed. Preprocess the data, including interpolation, rasterization and standardization, to ensure data consistency and accuracy.

[0182] Digital Elevation Model (DEM) Construction

[0183] Using remote sensing data and drone observation data, a digital elevation model of the watershed is constructed. DEM is the core of the virtual watershed, which provides detailed information on the basin topography and is used to simulate water flow paths and hydrodynamic processes within the basin.

[0184] Integration of watershed hydrological models

[0185] Combine physical models (such as rainfall-runoff models) and data-driven models (such as PINNs) to integrate watershed hydrological models in virtual watersheds. These models will be used to simulate hydrological processes such as rainfall, runoff, evaporation, and infiltration.

[0186] Virtual environment construction

[0187] Using computer graphics and virtual reality (VR) technology, the processed data and models are integrated into a three-dimensional virtual environment. Users can view the hydrological process at different time points and under different meteorological conditions in the virtual watershed, and even observe the impact on runoff by changing input conditions (such as rainfall and land use changes).

[0188] Experimental simulation and data calibration

[0189] The accuracy of the hydrological model can be calibrated and verified by conducting simulation experiments in a virtual watershed. For example, historical extreme weather events can be simulated and the simulation results can be compared with actual observational data to calibrate the model parameters.

[0190] Interactive analysis and decision support

[0191] Through the interactive tools provided by Virtual Watershed, users can dynamically adjust simulation parameters and observe changes in runoff processes. This interactivity makes Virtual Watershed a powerful decision support tool that helps users conduct predictive analysis and risk assessment in complex environments.

[0192] Feedback loops and model optimization

[0193] Collect user operation records and feedback information in the virtual watershed, combine with actual data, and continuously optimize the model and simulation accuracy of the virtual watershed. With the introduction of more real-time data, the prediction ability of the virtual watershed will gradually increase, becoming a system that continuously learns and evolves.

[0194] The beneficial effects of the present invention are:

[0195] This runoff prediction method based on multi-source data fusion, by fusing multi-source data and introducing a variety of deep learning models, not only enhances the model's ability to capture complex nonlinear relationships, but also ensures physical consistency and the ability to respond to extreme events; it provides more comprehensive runoff prediction results, helps to improve the robustness of the prediction and the reliability of the analysis results, overcomes the shortcomings of traditional models in terms of single data source, poor physical consistency, and insufficient prediction uncertainty analysis, and provides a more accurate and stable prediction tool for water resources management, disaster prevention and mitigation and other fields.

[0196] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0197] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A runoff prediction method based on multi-source data fusion, characterized in that: The following steps are involved: S1. Selection and collection of data sources Select multiple traditional data sources suitable for runoff prediction, including meteorological data, hydrological data, geographic data, and remote sensing data. At the same time, introduce non-traditional data sources, including social media data and crowd science data, to form multi-source data; and use drones for real-time observation to provide high-resolution ground information to enhance the understanding of terrain and vegetation cover; S2. Data preprocessing Preprocess multi-source data, including removing outliers, filling missing values, and standardizing data. For remote sensing data, image processing techniques such as cloud removal and noise filtering are used to improve data reliability due to factors such as cloud cover and sensor failure. S3. Design of data fusion strategy Design strategies suitable for multi-source data fusion, adopt a multi-level multi-source data fusion framework, and make comprehensive use of data from different sources based on the semantic level of the data during fusion; S4. Feature extraction and selection Extract relevant features from the fused data, including precipitation intensity, temperature change rate, surface vegetation index, and watershed area; and select and reduce the dimension of the extracted features to avoid feature redundancy and overfitting; S5. Model selection and construction Select appropriate prediction models according to the needs of runoff prediction, build a deep learning model of multi-model combination based on physical constraints, and introduce mixed reality technology to build a "virtual watershed". Combine simulation experiments with actual data to improve the prediction ability of the model; S6. Model training and validation The preprocessed data is input into the model for training, and two training models, a learning model and a reinforcement model, are used to evaluate the performance of the model. An adversarial training strategy is used to enhance the model's ability to predict extreme weather events when generating adversarial samples. S7. Model optimization and improvement After the initial training and validation of the model, the model is further optimized; S8. Runoff prediction and result analysis The optimized model is used to predict runoff in future time periods, and the accuracy and uncertainty of the prediction results are analyzed. The prediction results are combined with actual hydrological and meteorological conditions for reasonable interpretation and application to support water resources management and disaster prevention and mitigation work.

2. A runoff prediction method based on multi-source data fusion according to claim 1, characterized in that: In step S1: Among traditional data sources, meteorological data include precipitation, temperature, and humidity; hydrological data include river flow and water level; geographic data include topography and land use type; remote sensing data include satellite images and vegetation coverage; Among non-traditional data sources, social media data and crowd science data, real-time precipitation or water level change information in a specific area can be obtained by analyzing weather-related discussions, pictures and videos on social media; this can be used as supplementary information to traditional data sources, especially in areas where traditional data sources are lacking.

3. The runoff prediction method based on multi-source data fusion according to claim 1 is characterized in that: In step S2, the preprocessing of multi-source data is based on an adaptive data cleaning algorithm, which automatically selects the best data processing method according to different data sources and data types, including: Use autoencoder networks in deep learning to automatically detect and repair abnormal data; Generate reasonable filling of missing data through generative adversarial networks; A data source traceability system is established based on blockchain technology to verify the authenticity and reliability of the data.

4. The runoff prediction method based on multi-source data fusion according to claim 1 is characterized in that: In step S3, when the semantic level of data is considered during fusion, the following process is specifically included: First, low-level physical fusion of meteorological and hydrological data and medium-level physical fusion of social media and crowd science data are performed; Then, high-level feature fusion is performed in combination with geographic information data; Second, use graph neural networks to process complex geospatial relationships, or use self-attention mechanisms to capture long-distance dependencies between different data sources; Finally, by introducing federated learning technology, data privacy can be protected while data is shared between different organizations, further enriching the diversity of data sources.

5. The runoff prediction method based on multi-source data fusion according to claim 1 is characterized in that: In step S4, transfer learning and meta-learning techniques are also introduced to transfer learning feature extractors from similar hydrological prediction tasks, or use a meta-learning framework to quickly adjust the model to adapt to new data distribution; Image processing technology is used to extract complex spatial features from remote sensing data, or natural language processing technology is used to extract weather-related emotions and keywords from social media texts as auxiliary features.

6. The runoff prediction method based on multi-source data fusion according to claim 1 is characterized in that: In step S5, the construction of the deep learning model of the multi-model combination includes the following specific processes: Define the physical equations and select the physical equations suitable for the runoff process, including Saint-Venant equations, Navier-Stokes equations or rainfall-runoff model; Deep learning model design, building a deep neural network whose input is the features of multi-source data and output is runoff; Definition of loss function,The loss function consists of two parts: one is the traditional supervised learning loss, which is used to minimize the error between the model prediction value and the true observation value; the other is the physical constraint loss, which is used to ensure that the model prediction value satisfies the physical equation; the physical constraint loss is represented by the residual of the partial differential equation; Model training, using gradient descent method to train neural network parameters and correct the prediction results through physical equations; Model validation and tuning: evaluate model performance on the validation set to ensure that the model can not only accurately predict runoff but also meet the constraints of physical laws. Based on the validation results, adjust the network structure or optimization method to improve the model's predictive performance.

7. The runoff prediction method based on multi-source data fusion according to claim 1 is characterized in that: In step S6, the learning model and the reinforcement model respectively include the following contents: In the learning model, assume an input data set (X) that has no labels, design a learning task, and learn an effective representation (Z) of the input data by constructing a proxy task; Choose a proxy task, such as time series forecasting, spatiotemporal segment sorting, or masked data recovery tasks; For the time series data in runoff prediction (X = {x1, x2, ..., x T }), design a time segment prediction task: in, is the predicted data at the (t+k)th moment, (f θ ) is a deep learning model with parameters (θ); Use model (f θ ) is trained by minimizing the loss function of the proxy task (L SSL ), and learn an effective representation (Z) of the input data: Z=f θ (X) The loss function is defined as: Where (N) is the number of training samples, is the predicted value of the (i)th sample, is the corresponding true value; In the reinforcement model, in the framework of reinforcement learning, an environment (E) and an agent (A) are defined. The agent learns the optimal strategy (π(a|s)) by interacting with the environment, where (s) is the state and (a) is the action; The key variables in the runoff prediction task, such as precipitation, evaporation, and soil moisture, are taken as states (s t ): s t =[Precipitation t , evaporation t , soil moisture t , ...] The action chosen by the agent (a t ) is the adjustment of model parameters, feature selection or change of prediction strategy, adjusting the learning rate of the model or choosing to use a specific feature: a t =adjust(learning rate)or select(feature subset) Design a reward function (r t ), so that the agent chooses actions that maximize the accuracy of runoff prediction, for example, based on the negative of the prediction error as a reward: Among them, (y t ) is the actual runoff, is the predicted value of the model; Use the policy gradient method to optimize the strategy (π(a|s)) so that the cumulative reward maximize:

8. The runoff prediction method based on multi-source data fusion according to claim 1 is characterized in that: In step S7, the optimized contents include: Reselect or combine features and adjust the model structure, such as adding hidden layers and changing activation functions; Introducing new data sources, such as real-time remote sensing data, to improve the prediction accuracy of the model; An adaptive optimization algorithm is used to automatically adjust model parameters to improve the adaptability and predictive ability of the model, and the quantum optimization algorithm in quantum computing is used to accelerate the optimization process of the model and obtain better prediction results in a shorter time.

9. The runoff prediction method based on multi-source data fusion according to claim 1 is characterized in that: In step S8, in the runoff prediction, multiple possible runoff prediction scenarios are generated based on the deep learning model, and the confidence interval of each scenario is calculated; By modeling the probability distribution of network weights, multiple possible prediction results are generated. Assume that the output of the runoff prediction model is The input is (X t ), including precipitation, temperature, and terrain characteristics, the deep learning model is expressed as: Where (θ) is the weight of the neural network, is the training data set; Use variational inference or Markov Chain Monte Carlo methods to approximate the posterior distribution The loss function is defined as: Here, (q(θ)) is an approximation to the posterior distribution and (KL) is the Kullback-Leibler divergence, which measures the difference between two distributions.

10. The runoff prediction method based on multi-source data fusion according to claim 1 is characterized in that: In step S8, in the runoff prediction, the influence of different input variables on the prediction results is determined by a sensitivity analysis model, which helps to identify key environmental factors and potential risks; Assume that the output of runoff prediction is The input variables include multiple features It is used to evaluate the impact of each input variable on the output by calculating the partial derivative of the model output with respect to the input: Among them, (S i ) represents the input variable sensitivity to model output; By calculating the sensitivity of each input variable (S i ), rank the importance of variables and identify the most critical factors for runoff prediction for model optimization or result support.

Images