Reservoir multi-station water level joint prediction method based on federated learning

Abstract

The invention discloses a reservoir multi-station water level joint prediction method based on federal learning. The method comprises the following steps: S1, collecting and preprocessing water level data of multiple reservoir stations; s2, constructing a water level prediction model at each station, and performing independent training through a local gradient descent algorithm; s3, globally optimizing a local water level prediction result, and adjusting model parameters through an improved gradient weighted optimization algorithm; s4, uploading parameters of the water level prediction model to a central server through a federated learning mechanism, and performing safe aggregation on local water level prediction results by using a homomorphic encryption technology; s5, performing statistics on local water level prediction results after safe aggregation, and performing joint prediction on each reservoir site; and S6, dynamically adjusting the scheduling strategy of the reservoir, and constructing a station monitoring and early warning system to guide the scheduling operation. The method improves the water level prediction precision and the scheduling efficiency, guarantees the data privacy, and has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the field of federated learning technology, and in particular to a method for joint prediction of reservoir multi-station water levels based on federated learning. Background Art

[0002] With the global climate change and the growing demand for water resources management, reservoir scheduling and water level forecasting have become important issues in modern water management. Reservoir water level forecasting is of great significance for flood warning, drought management and rational allocation of water resources. Traditional water level forecasting methods mostly rely on data from a single reservoir for forecasting. Although these methods can provide reference information to a certain extent, the accuracy and stability of their forecast results are often limited due to the different geographical conditions, climate environments and hydrological characteristics between reservoirs. At the same time, the prediction model of a single reservoir is difficult to consider the mutual influence between multiple reservoirs, resulting in poor comprehensiveness and globality of water level forecasting. In order to improve the accuracy and adaptability of water level forecasting, researchers have gradually turned to water level joint forecasting methods based on multi-station data.

[0003] Traditional water level prediction methods typically rely on historical data from a single site, using techniques such as time series analysis and regression models to predict water levels. These methods mostly consider data changes at a single reservoir site, but ignore the interactions between multiple reservoirs and the impact of external environmental factors. In addition, traditional methods often face problems with data storage, processing efficiency, and privacy protection when processing large-scale water level data. Reservoir data typically includes historical water level data, real-time water level change trends, hydrological conditions, climate data, etc. Most of this data is distributed across different reservoir sites, and some of it is sensitive information. In reservoir scheduling and water level prediction, how to effectively integrate massive data from multiple sites and address issues such as privacy protection, data sharing, and communication efficiency has become a technical challenge that needs to be solved urgently.

[0004] To improve the accuracy and real-time performance of multi-site water level forecasts, distributed learning methods have been gradually introduced in recent years. In particular, the introduction of federated learning technology has provided new insights into collaborative forecasting across multiple reservoir sites. Federated learning uses distributed computing to enable multiple reservoir sites to train their own water level prediction models locally, while ensuring data privacy. This allows the model parameters to be aggregated and updated via a central server. This process eliminates the need to centrally store the raw data from each site, thus avoiding data leakage and privacy risks. However, despite the potential for federated learning in water level forecasting, its practical application still faces several challenges and technical bottlenecks.

[0005] First, while traditional federated learning mechanisms can ensure data privacy, their reliance on model parameter synchronization and updates often leads to inefficient communication. During the federated learning process, all participating sites need to upload their locally trained model parameters to a central server. The central server aggregates and updates the model parameters for each site, then distributes the updated global model parameters to each site. Due to data disparity and uneven computing power between reservoir sites, this process often requires frequent communication interactions, resulting in high network load and latency. Especially in the case of a large number of reservoir sites, communication efficiency becomes a key factor limiting overall system performance.

[0006] Secondly, traditional federated learning mechanisms ignore the spatial and hydrological interconnectedness between reservoir sites. When predicting water levels at multiple reservoirs, there are often geographical interactions between sites, such as the relationship between reservoir basins and the spatial distribution of precipitation. Existing federated learning methods primarily focus on training and aggregating local models, failing to effectively account for the differences in geographic, climatic, and hydrological conditions across reservoir sites, potentially resulting in inaccurate global forecasts.

[0007] Furthermore, current federated learning models mostly focus on privacy protection and independent training of local data, but lack the ability to optimize the global integration of multi-site data. Historical reservoir data, real-time water level trends, and hydrological environmental factors need to be rationally integrated through weighted fusion to improve prediction accuracy and global adaptability. Existing technologies often use a simple weighted averaging approach, ignoring the impact of different sites on global water level predictions. This results in insufficient accuracy and robustness in reservoir water level predictions.

[0008] In addition, when faced with large-scale, multi-source heterogeneous data, existing water level prediction models are often unable to handle complex data interactions and update processes. Especially in environments with large data volumes and poor communication conditions, the model training and update processes are inefficient, and the model performance is easily affected by the data quality of a single site.

[0009] Therefore, how to provide a method for joint prediction of reservoir multi-station water levels based on federated learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0010] One objective of the present invention is to propose a method for joint water-level prediction at multiple reservoir sites based on federated learning. This method leverages federated learning techniques, a weighted multi-site data fusion algorithm, and homomorphic encryption technology. It describes in detail how to achieve joint training and optimization of water-level prediction models for multiple reservoir sites through a federated learning mechanism while ensuring data privacy. Furthermore, the method uses an improved gradient-weighted optimization algorithm to fuse the water-level prediction results for each site, combining the correlations between reservoirs, their geographic locations, and hydrological conditions to generate an accurate global water-level prediction.

[0011] A method for joint prediction of reservoir multi-station water levels based on federated learning according to an embodiment of the present invention includes the following steps:

[0012] S1. Collect water level data from multiple reservoir sites and perform preprocessing;

[0013] S2. Based on the preprocessed water level data, a water level prediction model is constructed at each reservoir site, and independently trained using a local gradient descent algorithm to output the local water level prediction results for each reservoir site;

[0014] S3. Globally optimize the local water level prediction results and dynamically adjust the water level prediction model parameters using an improved gradient weighted optimization algorithm that performs weighted updates based on historical data from each site and real-time water level change trends;

[0015] S4. Upload the water level prediction model parameters to the central server through a federated learning mechanism, and securely aggregate the local water level prediction results using homomorphic encryption technology. The federated learning mechanism uses asynchronous updates and differential privacy protection technology to optimize network communication efficiency.

[0016] S5. Calculate the local water level prediction results after security aggregation and perform a joint prediction for each reservoir site. The joint prediction generates a global water level prediction result by integrating the water level prediction results of multiple sites and combining the correlation, geographical location and hydrological conditions between the sites.

[0017] S6. Dynamically adjust the reservoir scheduling strategy based on the global water level prediction results, build a site monitoring and early warning system, and guide reservoir scheduling operations based on monitoring and early warning signals.

[0018] Optionally, the water level data includes real-time water level data, historical water level data, water level anomaly data and reservoir scheduling historical data.

[0019] Optionally, the preprocessing includes deduplication, missing value filling and outlier processing.

[0020] Optionally, the scheduling strategy includes water level regulation, flood discharge operation and water storage management, and the site monitoring and early warning system is used to monitor the dynamics of reservoir water levels, detect abnormal water level changes in a timely manner, and issue early warning signals.

[0021] Optionally, the S2 specifically includes:

[0022] S21, based on the pre-processed water level data, for any reservoir site i water level data D i , build water level prediction model M i , where D i ={w i,1 ,w i,2 ,…,w i,n} represents the water level data sequence of site i, w i,k represents the water level value of site i at time point k, and n is the length of the water level data sequence;

[0023] S22, water level data D at reservoir site i i The parameters θ of the water level prediction model are optimized by the local gradient descent algorithm. i , which enables the model to accurately predict water levels within a given time period:

[0024]

[0025] Among them, L i (θ i ) is the loss function, For the model at time point k to w i,k The predicted value of represents the local water level prediction result;

[0026] S23, calculate the gradient value corresponding to each site i to optimize the water level prediction model parameters, and use the local gradient descent algorithm to optimize the model parameters θ i To update:

[0027]

[0028] Among them, η is the learning rate, which controls the step size of each update. is the gradient value corresponding to site i, is the gradient value corresponding to the local water level prediction result;

[0029] S24. After completing independent training, the water level prediction model of each site i outputs the local water level prediction results. The local prediction results include water level prediction information of each reservoir site at each time point.

[0030] Optionally, the water level prediction model is independently trained on the water level data of each reservoir site through a local gradient descent algorithm, and the local water level prediction results of each reservoir site are calculated. The training process minimizes the loss function of the prediction error and uses the calculated gradient to update the model parameters, and finally outputs the optimized water level prediction model and corresponding local water level prediction results for each reservoir site.

[0031] Optionally, the S3 specifically includes:

[0032] S31, based on the local water level prediction results of each reservoir site i And the historical water level data H of each station i ={h i,1 ,h i,2 ,…,h i,m} and real-time water level change trend T i ={t i,1 ,t i,2 ,…,t i,p}, perform weighted update on the local water level prediction results of each station and calculate the weighting factor:

[0033]

[0034] Among them, h i,j represents the historical water level data of site i at time point j, t i,k is the real-time water level change trend of site i at time point k, m and p are the lengths of historical water level data and real-time water level change trend respectively, α i is the weighting factor based on historical water level data, β i is the weighting factor based on the real-time water level change trend, w i,j and w i,k are the weight coefficients of historical water level data and real-time water level change trend, h i,j and t i,k They are the corresponding historical water level data and real-time water level change trend values;

[0035] S32, based on weighting factor α i and β i , using the improved gradient weighted optimization algorithm, the water level prediction model parameters θ of each reservoir site i are updated twice i :

[0036]

[0037] Among them, Z i (θ i ) is the objective function, L i (θ i ) is the loss function, and are the historical water level data and real-time water level change trend predicted by the water level prediction model, respectively. The second and third terms in the objective function are the weighted losses of the historical water level data and real-time water level change trend, respectively.

[0038] Optionally, the improved gradient weighted optimization algorithm introduces a weighting factor α based on historical water level data and real-time water level change trends. i and β i , dynamically adjust the weights during the local gradient descent process, perform weighted optimization on the gradient according to the spatiotemporal correlation between sites, and update the model parameters θ twice i , and accelerate the convergence process through weighted gradient update.

[0039] Optionally, the S4 specifically includes:

[0040] S41, through the federated learning mechanism, the water level prediction model parameters θ of each reservoir site i are i Uploaded to the central server, the parameter θ i Contains the weight information learned by each site during the local training process;

[0041] S42. After the central server receives the model parameters uploaded by each reservoir site, it uses homomorphic encryption technology to encrypt the water level prediction results of each reservoir site. The encrypted results are used for secure aggregation to ensure that data privacy is not leaked during transmission;

[0042] S43. Gradually updating the water level prediction model of each reservoir site through an asynchronous update mechanism. The asynchronous update mechanism allows each site to update independently without synchronization, avoiding the time delay of waiting for all sites to complete the update, thereby improving network communication efficiency and overall computing efficiency;

[0043] S44. After receiving the water level prediction model parameters from the central server, each reservoir site continues to perform local training based on the local water level data. The local training process is based on the encryption parameters of the central server to ensure that the privacy of the data is not leaked during the local training.

[0044] S45. When uploading local water level prediction results at reservoir sites, differential privacy protection technology is applied to the uploaded data to ensure that the private data of each site cannot be inferred or reversed by adding noise;

[0045] S46. By combining homomorphic encryption, asynchronous update and differential privacy protection technologies, network communication efficiency is optimized, the security of site data transmission is improved, and the data privacy of reservoir sites is effectively protected.

[0046] Optionally, the S5 specifically includes:

[0047] S51. After receiving the local water level prediction results protected by homomorphic encryption and differential privacy, the central server performs decryption processing to restore the local water level prediction values of each reservoir site and ensure data security;

[0048] S52, statistically summarizing the local water level prediction results of each reservoir site, and performing weighted processing based on the geographical location, historical water level data, real-time water level change trend and hydrological conditions of each site to calculate a weighted coefficient for each site;

[0049] S53. Based on the weighting coefficients, the local water level prediction results of multiple reservoir sites are weightedly fused. The weighted fusion generates a global water level prediction result by considering the spatial correlation between the sites, the correlation of historical data and the hydrological conditions.

[0050] The beneficial effects of the present invention are:

[0051] First, this invention addresses the data privacy issues inherent in traditional water level prediction methods by employing a federated learning mechanism. When jointly predicting water level data from multiple reservoir sites, this invention eliminates the need to centralize the raw data from each site, ensuring data privacy. By leveraging homomorphic encryption and differential privacy protection technologies, this invention effectively mitigates the risk of data leakage and provides a privacy-preserving prediction framework for reservoir sites.

[0052] Secondly, the present invention uses an improved gradient weighted optimization algorithm to fully consider the correlations, geographical locations, and hydrological differences among multiple reservoir sites during the global optimization process, thereby improving the accuracy of prediction results. By weightedly integrating the water level prediction results of each site, the present invention generates a global water level prediction with higher accuracy and stability, better adapting to the actual conditions of different reservoirs and effectively improving the decision-making basis for reservoir scheduling and management.

[0053] Finally, this invention optimizes network communication efficiency through an asynchronous update mechanism. In traditional federated learning mechanisms, frequent communication interactions often lead to network delays and wasted computing resources. However, this asynchronous update strategy significantly reduces the frequency and latency of data transmission, improving overall system efficiency and making it particularly suitable for collaborative predictions at large-scale reservoir sites. This method effectively enhances the real-time prediction capabilities of large-scale reservoir networks, providing faster and more reliable decision support for reservoir management. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0055] Figure 1 This is a flow chart of a method for joint prediction of reservoir water levels at multiple stations based on federated learning proposed by the present invention;

[0056] Figure 2 This is a schematic diagram of the local water level prediction process of a reservoir multi-station water level joint prediction method based on federated learning proposed by the present invention;

[0057] Figure 3 This is a schematic diagram of the entire water level prediction process of a method for joint prediction of multi-station water levels in a reservoir based on federated learning proposed in the present invention. DETAILED DESCRIPTION

[0058] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0059] refer to Figure 1-3 A method for joint prediction of reservoir water levels at multiple stations based on federated learning includes the following steps:

[0060] S1. Collect water level data from multiple reservoir sites and perform preprocessing;

[0061] S2. Based on the preprocessed water level data, a water level prediction model is constructed at each reservoir site, and independently trained using a local gradient descent algorithm to output the local water level prediction results for each reservoir site;

[0062] S3. Globally optimize the local water level prediction results and dynamically adjust the water level prediction model parameters using an improved gradient weighted optimization algorithm that performs weighted updates based on historical data from each site and real-time water level change trends;

[0063] S4. Upload the water level prediction model parameters to the central server through a federated learning mechanism, and securely aggregate the local water level prediction results using homomorphic encryption technology. The federated learning mechanism uses asynchronous updates and differential privacy protection technology to optimize network communication efficiency.

[0064] S5. Calculate the local water level prediction results after security aggregation and perform a joint prediction for each reservoir site. The joint prediction generates a global water level prediction result by integrating the water level prediction results of multiple sites and combining the correlation, geographical location and hydrological conditions between the sites.

[0065] S6. Dynamically adjust the reservoir scheduling strategy based on the global water level prediction results, build a site monitoring and early warning system, and guide reservoir scheduling operations based on monitoring and early warning signals.

[0066] This paper combines federated learning with an improved gradient-weighted optimization algorithm to achieve joint water level forecasting at multiple reservoir sites, ensuring data privacy while improving prediction accuracy. Through local training and global optimization, it generates accurate global water level forecasts, taking into account geographic correlations and real-time trends. Using homomorphic encryption, asynchronous updates, and differential privacy protection, it optimizes communication efficiency and data security. Ultimately, the water level forecast results are used to dynamically adjust reservoir scheduling strategies and provide real-time decision support.

[0067] In this embodiment, the water level data includes real-time water level data, historical water level data, water level anomaly data and reservoir scheduling historical data.

[0068] The present invention ensures the comprehensiveness and representativeness of input data by comprehensively integrating different types of water level data, including real-time water level data, historical water level data, abnormal data and reservoir scheduling history data.

[0069] In this embodiment, the preprocessing includes deduplication, missing value filling and outlier processing.

[0070] The present invention uses duplicate removal, missing value filling, and outlier handling in data preprocessing to ensure data quality and integrity. This refined data processing approach avoids the impact of noisy data on water level prediction results and improves the stability and robustness of the entire model.

[0071] In this embodiment, the dispatching strategy includes water level regulation, flood discharge operation and water storage management. The site monitoring and early warning system is used to monitor the dynamics of reservoir water levels, detect abnormal water level changes in a timely manner, and issue early warning signals.

[0072] This invention dynamically adjusts reservoir scheduling strategies based on global water level forecasts. By building a site-level monitoring and early warning system, it can promptly detect abnormal water level changes and issue warning signals. This functionality not only strengthens reservoir monitoring capabilities but also provides a more accurate and real-time basis for reservoir scheduling decisions, ensuring the safety and efficiency of reservoir operation.

[0073] In this embodiment, S2 specifically includes:

[0074] S21, based on the pre-processed water level data, for any reservoir site i water level data D i , build water level prediction model M i , where D i ={w i,1 ,w i,2 ,…,w i,n} represents the water level data sequence of site i, w i,k represents the water level value of site i at time point k, and n is the length of the water level data sequence;

[0075] S22, water level data D at reservoir site i i The parameters θ of the water level prediction model are optimized by the local gradient descent algorithm. i , which enables the model to accurately predict water levels within a given time period:

[0076]

[0077] Among them, L i (θ i ) is the loss function, For the model at time point k to w i,k The predicted value of represents the local water level prediction result;

[0078] S23, calculate the gradient value corresponding to each site i to optimize the water level prediction model parameters, and use the local gradient descent algorithm to optimize the model parameters θ i To update:

[0079]

[0080] Among them, η is the learning rate, which controls the step size of each update. is the gradient value corresponding to site i, is the gradient value corresponding to the local water level prediction result;

[0081] S24. After completing independent training, the water level prediction model of each site i outputs the local water level prediction results. The local prediction results include water level prediction information of each reservoir site at each time point.

[0082] This method ensures model accuracy at each reservoir site by independently constructing a water level prediction model and optimizing it using a local gradient descent algorithm. Furthermore, dynamic updates based on historical water level data and real-time water level trends at each site ensure the real-time and adaptability of model parameters. Ultimately, the local prediction results output by each site provide accurate foundational data for subsequent global optimization.

[0083] In this embodiment, the water level prediction model is independently trained on the water level data of each reservoir site through a local gradient descent algorithm, and the local water level prediction results of each reservoir site are calculated. The training process minimizes the loss function of the prediction error and uses the calculated gradient to update the model parameters, and finally outputs the optimized water level prediction model and corresponding local water level prediction results for each reservoir site.

[0084] The present invention ensures the prediction accuracy of each site by independently training the water level prediction model at each reservoir site, optimizing the model parameters using a local gradient descent algorithm, minimizing the prediction error, and providing an optimized water level prediction model, thereby improving the accuracy and reliability of the overall prediction.

[0085] In this embodiment, S3 specifically includes:

[0086] S31, based on the local water level prediction results of each reservoir site i And the historical water level data H of each station i ={h i,1 ,h i,2 ,…,h i,m} and real-time water level change trend T i ={t i,1 ,t i,2 ,…,t i,p}, perform weighted update on the local water level prediction results of each station and calculate the weighting factor:

[0087]

[0088] Among them, h i,j represents the historical water level data of site i at time point j, t i,k is the real-time water level change trend of site i at time point k, m and p are the lengths of historical water level data and real-time water level change trend respectively, α i is the weighting factor based on historical water level data, β i is the weighting factor based on the real-time water level change trend, w i,j and w i,k are the weight coefficients of historical water level data and real-time water level change trend, h i,j and t i,k They are the corresponding historical water level data and real-time water level change trend values;

[0089] S32, based on weighting factor α i and β i , using the improved gradient weighted optimization algorithm, the water level prediction model parameters θ of each reservoir site i are updated twice i :

[0090]

[0091] Among them, Z i (θ i ) is the objective function, L i (θ i ) is the loss function, and are the historical water level data and real-time water level change trend predicted by the water level prediction model, respectively. The second and third terms in the objective function are the weighted losses of the historical water level data and real-time water level change trend, respectively.

[0092] This paper uses an improved gradient-weighted optimization algorithm to perform weighted updates based on each site's local water level forecast results, historical water level data, and real-time change trends. This optimization algorithm not only improves the efficiency of updating model parameters at each site but also enables efficient secondary updates of model parameters by dynamically adjusting weighting factors, further improving the accuracy of global water level forecasts.

[0093] In this embodiment, the improved gradient weighted optimization algorithm introduces a weighting factor α based on historical water level data and real-time water level change trend. i and β i , dynamically adjust the weights during the local gradient descent process, perform weighted optimization on the gradient according to the spatiotemporal correlation between sites, and update the model parameters θ twice i , and accelerate the convergence process through weighted gradient update.

[0094] This implementation introduces weighting factors based on historical water level data and real-time water level change trends, dynamically adjusts weights during the local gradient descent process, optimizes gradients and accelerates model convergence, thereby improving the accuracy and stability of water level prediction.

[0095] In this embodiment, the S4 specifically includes:

[0096] S41, through the federated learning mechanism, the water level prediction model parameters θ of each reservoir site i are i Uploaded to the central server, the parameter θ i Contains the weight information learned by each site during the local training process;

[0097] S42. After the central server receives the model parameters uploaded by each reservoir site, it uses homomorphic encryption technology to encrypt the water level prediction results of each reservoir site. The encrypted results are used for secure aggregation to ensure that data privacy is not leaked during transmission;

[0098] S43. Gradually updating the water level prediction model of each reservoir site through an asynchronous update mechanism. The asynchronous update mechanism allows each site to update independently without synchronization, avoiding the time delay of waiting for all sites to complete the update, thereby improving network communication efficiency and overall computing efficiency;

[0099] S44. After receiving the water level prediction model parameters from the central server, each reservoir site continues to perform local training based on the local water level data. The local training process is based on the encryption parameters of the central server to ensure that the privacy of the data is not leaked during the local training.

[0100] S45. When uploading local water level prediction results at reservoir sites, differential privacy protection technology is applied to the uploaded data to ensure that the private data of each site cannot be inferred or reversed by adding noise;

[0101] S46. By combining homomorphic encryption, asynchronous update and differential privacy protection technologies, network communication efficiency is optimized, the security of site data transmission is improved, and the data privacy of reservoir sites is effectively protected.

[0102] This invention protects the privacy of water level data through homomorphic encryption technology, adopts an asynchronous update mechanism to reduce communication delays and improve data transmission efficiency, and uses differential privacy protection technology to prevent data leakage and reverse inference, thereby optimizing privacy protection and communication efficiency and providing technical support for large-scale multi-site collaborative prediction.

[0103] In this embodiment, the S5 specifically includes:

[0104] S51. After receiving the local water level prediction results protected by homomorphic encryption and differential privacy, the central server performs decryption processing to restore the local water level prediction values of each reservoir site and ensure data security;

[0105] S52, statistically summarizing the local water level prediction results of each reservoir site, and performing weighted processing based on the geographical location, historical water level data, real-time water level change trend and hydrological conditions of each site to calculate a weighted coefficient for each site;

[0106] S53. Based on the weighting coefficients, the local water level prediction results of multiple reservoir sites are weightedly fused. The weighted fusion generates a global water level prediction result by considering the spatial correlation between the sites, the correlation of historical data and the hydrological conditions.

[0107] The present invention aggregates and weightedly fuses the local water level prediction results of multiple reservoir sites through a central server, combines geographical location, historical water level data, real-time change trends and hydrological conditions, and generates global water level prediction results, thereby optimizing reservoir scheduling and management decisions.

[0108] Example 1:

[0109] In order to verify the feasibility of the present invention in implementation, the present invention is applied to multiple reservoir sites in a certain area to conduct multi-site joint prediction of water level forecasts. The geographical conditions of the reservoirs in this area are complex, and the water level fluctuates greatly. The traditional single-site prediction method often cannot accurately reflect the water level change trend of the entire region, and lacks consideration of the water level correlation between each site. Therefore, the present invention adopts a multi-site water level joint prediction method based on federated learning, aiming to improve the accuracy and reliability of water level prediction through data privacy protection, global optimization and weighted fusion technology, thereby providing a scientific decision-making basis for the scheduling and management of reservoirs.

[0110] In this example, a total of 10 reservoir sites are involved, distributed across diverse geographical locations and hydrological environments. Water level data for each site includes historical water level data, real-time water level trends, and historical reservoir operation data. By collecting and preprocessing this data, including deduplication, missing value filling, and outlier handling, high-quality data input is provided for model training. Each site is trained independently locally, using a local gradient descent algorithm to optimize the water level prediction model and output local water level prediction results for each site.

[0111] The data training and prediction process utilizes an improved gradient-weighted optimization algorithm. This algorithm dynamically adjusts the water level prediction model parameters for each site by factoring in historical water level data and real-time water level trends. After local model training is complete, model parameters for all reservoir sites are uploaded to a central server for aggregation via a federated learning mechanism. Homomorphic encryption and differential privacy technologies ensure data security during upload and aggregation.

[0112] After aggregating model parameters from each site, the central server uses weighted fusion technology to generate a global water level forecast based on each site's geographic location, historical water level data, real-time water level trends, and hydrological conditions. This global forecast not only integrates water level forecast information from each site but also weights it to account for the actual conditions at each reservoir site, resulting in a more accurate global water level forecast.

[0113] In order to evaluate the effect of the present invention, the traditional single-site prediction method and the multi-site joint prediction method based on federated learning proposed by the present invention were compared in the experiment. Experimental data show that the global water level prediction results generated by the method of the present invention have significantly reduced prediction errors compared with the traditional single-site prediction method, especially when the water level fluctuates, the prediction accuracy has been significantly improved. For example, over a period of time (from June 1, 2019 to June 30, 2019), the error of the traditional prediction method at a certain reservoir site was 8.5 cm, while the error after adopting the method of the present invention was only 3.2 cm, a reduction of 62%. The following is a table of some data from the experiment:

[0114] Table 1 Comparison results of water level prediction at reservoir sites

[0115]

[0116]

[0117] As can be seen in Table 1, the water level prediction errors using the present method are generally lower than those of traditional methods. In particular, the errors remain relatively low across multiple predictions, demonstrating the significant advantages of the present method in multi-site joint prediction. By weighted fusion of data from multiple reservoir sites, combined with the reservoir's geographic location, historical water levels, real-time change trends, and hydrological conditions, the accuracy of the global prediction results is effectively improved.

[0118] In this embodiment, the present invention not only addresses the inability of single-site prediction methods to fully utilize multi-site information, but also effectively improves the accuracy of the prediction model through a federated learning mechanism while ensuring data privacy. Furthermore, experimental results demonstrate that the present invention can significantly optimize reservoir scheduling decisions, improving the scientific nature and timeliness of reservoir management, and has broad application prospects.

[0119] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for joint prediction of reservoir water levels at multiple stations based on federated learning, characterized in that: The steps include: S1. Collect water level data from multiple reservoir sites and perform preprocessing; S2. Based on the preprocessed water level data, a water level prediction model is constructed at each reservoir site, and independently trained using a local gradient descent algorithm to output the local water level prediction results for each reservoir site; S3. Globally optimize the local water level prediction results and dynamically adjust the water level prediction model parameters using an improved gradient weighted optimization algorithm that performs weighted updates based on historical data from each site and real-time water level change trends; S4. Upload the water level prediction model parameters to the central server through a federated learning mechanism, and securely aggregate the local water level prediction results using homomorphic encryption technology. The federated learning mechanism uses asynchronous updates and differential privacy protection technology to optimize network communication efficiency. S5. Calculate the local water level prediction results after security aggregation and perform a joint prediction for each reservoir site. The joint prediction generates a global water level prediction result by integrating the water level prediction results of multiple sites and combining the correlation, geographical location and hydrological conditions between the sites. S6. Dynamically adjust the reservoir scheduling strategy based on the global water level prediction results, build a site monitoring and early warning system, and guide reservoir scheduling operations based on monitoring and early warning signals.

2. The method for joint prediction of reservoir water levels at multiple stations based on federated learning according to claim 1 is characterized in that: The water level data includes real-time water level data, historical water level data, water level anomaly data and reservoir scheduling historical data.

3. The method for joint prediction of reservoir water levels at multiple stations based on federated learning according to claim 1 is characterized in that: The preprocessing includes deduplication, missing value filling and outlier processing.

4. The method for joint prediction of reservoir water levels at multiple stations based on federated learning according to claim 1 is characterized in that: The dispatching strategy includes water level regulation, flood discharge operation and water storage management. The site monitoring and early warning system is used to monitor the dynamics of reservoir water levels, detect abnormal water level changes in a timely manner, and issue early warning signals.

5. The method for joint prediction of reservoir water levels at multiple stations based on federated learning according to claim 1 is characterized in that: The S2 specifically includes: S21, based on the pre-processed water level data, for any reservoir site i water level data D i , build water level prediction model M i , where D i ={w i,1 ,w i,2 ,…,w i,n } represents the historical water level data sequence of site i, w i,k represents the water level value of site i at time point k, and n is the length of the water level data sequence; S22, water level data D at reservoir site i i The parameters θ of the water level prediction model are optimized by the local gradient descent algorithm. i , which enables the model to accurately predict water levels within a given time period: Among them, L i (θ i ) is the loss function, For the model at time point k to w i,k The predicted value of represents the local water level prediction result; S23, calculate the gradient value corresponding to each site i to optimize the water level prediction model parameters, and use the local gradient descent algorithm to optimize the model parameters θ i To update: Among them, η is the learning rate, which controls the step size of each update. is the gradient value corresponding to site i, is the gradient value corresponding to the local water level prediction result; S24. After completing independent training, the water level prediction model of each site i outputs the local water level prediction results. The local prediction results include water level prediction information of each reservoir site at each time point.

6. The method for joint prediction of reservoir water levels at multiple stations based on federated learning according to claim 5 is characterized in that: The water level prediction model is independently trained on the historical water level data of each reservoir site using a local gradient descent algorithm, and the local water level prediction results of each reservoir site are calculated. The training process minimizes the loss function of the prediction error and uses the calculated gradient to update the model parameters, ultimately outputting the optimized water level prediction model and the corresponding local water level prediction results for each reservoir site.

7. The method for joint prediction of reservoir multi-station water levels based on federated learning according to claim 1 is characterized in that: The S3 specifically includes: S31, based on the local water level prediction results of each reservoir site i And the historical water level data H of each station i ={h i,1 ,h i,2 ,…,h i,m } and real-time water level change trend T i ={t i,1 ,t i,2 ,…,t i,p }, perform weighted update on the local water level prediction results of each station and calculate the weighting factor: Among them, h i,j represents the historical water level data of site i at time point j, t i,k is the real-time water level change trend of site i at time point k, m and p are the lengths of historical water level data and real-time water level change trend respectively, α i is the weighting factor based on historical water level data, β i is the weighting factor based on the real-time water level change trend, w i,j and w i,k are the weight coefficients of historical water level data and real-time water level change trend, h i,j and t i,k They are the corresponding historical water level data and real-time water level change trend values; S32, based on weighting factor α i and β i , using the improved gradient weighted optimization algorithm, the water level prediction model parameters θ of each reservoir site i are updated twice i : Among them, Z i (θ i ) is the objective function, L i (θ i ) is the loss function, and are the historical water level data and real-time water level change trend predicted by the water level prediction model, respectively. The second and third terms in the objective function are the weighted losses of the historical water level data and real-time water level change trend, respectively.

8. The method for joint prediction of reservoir water levels at multiple stations based on federated learning according to claim 7 is characterized in that: The improved gradient weighted optimization algorithm introduces a weighting factor α based on historical water level data and real-time water level change trend. i and β i , dynamically adjust the weights during the local gradient descent process, perform weighted optimization on the gradient according to the spatiotemporal correlation between sites, and update the model parameters θ twice i , and accelerate the convergence process through weighted gradient update.

9. The method for joint prediction of reservoir water levels at multiple stations based on federated learning according to claim 1, characterized in that: The S4 specifically includes: S41, through the federated learning mechanism, the water level prediction model parameters θ of each reservoir site i are i Uploaded to the central server, the parameter θ i Contains the weight information learned by each site during the local training process; S42. After the central server receives the model parameters uploaded by each reservoir site, it uses homomorphic encryption technology to encrypt the water level prediction results of each reservoir site. The encrypted results are used for secure aggregation to ensure that data privacy is not leaked during transmission; S43. Gradually updating the water level prediction model of each reservoir site through an asynchronous update mechanism. The asynchronous update mechanism allows each site to update independently without synchronization, avoiding the time delay of waiting for all sites to complete the update, thereby improving network communication efficiency and overall computing efficiency; S44. After receiving the water level prediction model parameters from the central server, each reservoir site continues to perform local training based on the local water level data. The local training process is based on the encryption parameters of the central server to ensure that the privacy of the data is not leaked during the local training. S45. When uploading local water level prediction results at reservoir sites, differential privacy protection technology is applied to the uploaded data to ensure that the private data of each site cannot be inferred or reversed by adding noise; S46. By combining homomorphic encryption, asynchronous update and differential privacy protection technologies, network communication efficiency is optimized, the security of site data transmission is improved, and the data privacy of reservoir sites is effectively protected.

10. The method for joint prediction of reservoir water levels at multiple stations based on federated learning according to claim 1, characterized in that: The S5 specifically includes: S51. After receiving the local water level prediction results protected by homomorphic encryption and differential privacy, the central server performs decryption processing to restore the local water level prediction values of each reservoir site and ensure data security; S52, statistically summarizing the local water level prediction results of each reservoir site, and performing weighted processing based on the geographical location, historical water level data, real-time water level change trend and hydrological conditions of each site to calculate a weighted coefficient for each site; S53. Based on the weighting coefficients, the local water level prediction results of multiple reservoir sites are weightedly fused. The weighted fusion generates a global water level prediction result by considering the spatial correlation between the sites, the correlation of historical data and the hydrological conditions.

Images