Hydrological sequence missing data complementation and trend prediction system and method

Abstract

The invention discloses a hydrological sequence missing data complementation and trend prediction system and method, and belongs to the technical field of water conservancy information. The system comprises a data acquisition and preprocessing module, a spatial-temporal feature fusion module, a physical constraint completion module, an uncertainty prediction module, a credible evidence storage and verification module and a visual tracing module which are connected in sequence. The method comprises the following steps: collecting and preprocessing multi-source hydrological data; combining the spatial-temporal features through a dual-channel network to generate joint representation; generating and checking a completion sequence based on joint representation and coupling physical constraints such as a unit line method and a Manning formula; performing uncertainty trend prediction by using an integrated predictor; the key data of the whole process are linked and stored, and are automatically verified through an intelligent contract; and full-life-cycle visual tracing is provided. According to the method, the complementation robustness and accuracy in a high-missing-rate scene are effectively improved, the physical consistency of a complementation result is ensured, and credible evidence storage and transparent decision of the whole process are realized.

Description

A system and method for completing missing hydrological series data and predicting trends. Technical Field

[0001] This invention belongs to the field of water conservancy information technology, specifically relating to a system and method for completing missing hydrological sequence data and predicting trends. Background Technology

[0002] With the deepening of the construction of smart water conservancy systems, hydrological sequence data, as the core foundation for refined water resource management, accurate early warning of floods and droughts, and assessment of watershed ecological health, has become a key support for modern water governance systems in terms of its completeness, timeliness, and reliability. Wide-area distributed hydrological sensing networks built upon Internet of Things (IoT) technology have achieved high-frequency, multi-source synchronous acquisition of key hydrological elements such as water level, flow, and rainfall, significantly improving the spatiotemporal coverage of hydrological monitoring. However, in complex field environments, factors such as aging sensor hardware, interrupted wireless communication links, extreme weather interference, or power supply system fluctuations lead to widespread random, patchy, or even continuous data gaps lasting several days in the acquired time series data. Such data deficiencies not only disrupt the inherent temporal continuity and statistical stationarity of hydrological processes but also severely restrict the reliability and decision-making effectiveness of hydrological models in flood evolution simulation, drought trend assessment, and long-term climate change response analysis.

[0003] Among these, hydrological data completion and trend prediction technologies, which restore missing information and predict future evolution trends through intelligent modeling, are core components for improving the usability of hydrological data. Existing technologies mainly rely on traditional interpolation methods (such as linear interpolation and cubic splines) or shallow machine learning models (such as support vector regression and random forests) for missing value estimation. While these methods perform reasonably well in scenarios with low missing rates (<20%), they struggle to depict the highly nonlinear and dynamic evolutionary patterns of hydrological systems. In recent years, deep learning-based temporal modeling methods (such as LSTM and GRU) have been introduced into hydrological data completion tasks due to their powerful contextual modeling capabilities. However, they generally employ a single-site isolated modeling paradigm, neglecting the close spatiotemporal coordination between upstream and downstream stations, and between main and tributary stations, driven by meteorological and hydrodynamic coupling. When a target station encounters a high proportion (>40%) or continuous missing data, relying solely on its own historical sequences can easily lead to a severe lack of information entropy, resulting in completion results that often exhibit physical distortions (such as water level jumps violating mass conservation) or logical contradictions (such as downstream water levels responding to rainfall before upstream levels). Even though some solutions attempt to introduce data from neighboring sites as auxiliary input, they lack a dynamic verification mechanism for the credibility of the reference data source. Once the data from neighboring sites is abnormal or maliciously tampered with, the erroneous information will be amplified and spread through the model structure, further weakening the overall robustness of the system.

[0004] Existing technologies generally suffer from centralized architecture flaws, lacking verifiable and tamper-proof governance mechanisms across the entire data collection, transmission, storage, and computation chain. In cross-departmental and cross-regional collaborative application scenarios, data silos and trust barriers are prominent issues. Even with high-precision algorithm outputs, the results are difficult to gain acceptance from business stakeholders due to unclear sources and opaque processes. Furthermore, the black-box nature of deep neural networks presents a profound contradiction with the rigid requirements of water conservancy management for decision interpretability and process auditability: existing solutions cannot transparently record key intermediate variables such as physical constraint verification and spatial attention weight allocation during the completion process, nor can they support complete retrospective analysis of the prediction conclusion generation path. Therefore, there is an urgent need to construct a new intelligent hydrological analysis framework that deeply integrates artificial intelligence, blockchain, and IoT technologies. This framework should systematically overcome the collaborative bottlenecks in three dimensions: robustness against high missing rates, data credibility, and system auditability by embedding physical laws, constructing spatiotemporal collaborative models, and implementing full-process trusted evidence storage. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a system and method for completing missing hydrological sequence data and predicting trends.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a system for completing missing hydrological sequence data and predicting trends, comprising:

[0008] The data acquisition and preprocessing module is used to acquire multi-source hydrological time-series data and perform edge-side preprocessing.

[0009] The spatiotemporal feature fusion module is used to extract the temporal features of the target station and fuse the spatial information of neighboring stations to generate a joint spatiotemporal representation.

[0010] The physical constraint completion module is used to generate a completion sequence based on the joint spatiotemporal representation and to verify and correct it using hydrophysical constraints.

[0011] The uncertainty prediction module is used to perform integrated predictions based on the completed sequence and output the prediction results and their confidence intervals.

[0012] The trusted evidence storage and verification module is used to store key system process data on the blockchain and perform automated verification through smart contract coding; and

[0013] The visualization traceability module is used to provide full lifecycle data lineage tracing and visualization from data collection to predictive output.

[0014] Optionally, the spatiotemporal feature fusion module includes:

[0015] The local feature extraction unit uses a recurrent neural network to extract the temporal dependency features of the target site;

[0016] The spatial correlation fusion unit models the hydrological correlation between stations based on the structure of dynamic graphs and aggregates information from neighboring stations through an attention mechanism.

[0017] The edge weights in the dynamic graph are determined based on the statistical correlation of data between stations and the physical delay of water flow propagation.

[0018] Optionally, in the spatial association fusion unit, the edge weights Calculated dynamically using the following formula:

[0019]

[0020] Where t represents the center time of the current processing time window, Indicates site and Pearson correlation coefficient within the sliding time window This represents the estimated water propagation time based on the river channel distance and average flow velocity between two stations. This is the empirical decay constant.

[0021] Optionally, the physical constraint completion module includes:

[0022] A sequence generation unit is used to generate a preliminary completed sequence based on the joint spatiotemporal representation;

[0023] The physical verification unit integrates unit line method flow constraint and / or Manning formula flow velocity constraint, and is used to perform real-time rationality verification on the preliminary completed sequence.

[0024] The correction unit is used to trigger the sequence generation unit to regenerate the candidate completion sequence when the physical verification unit determines that there is an anomaly, until the output completion sequence meets the preset physical verification rules.

[0025] Optionally, the physical verification unit may preset verification rules including at least one of the following:

[0026] The rate of change of water level in the completed sequence exceeds a threshold determined based on historical data;

[0027] The watershed confluence time inferred from the completed sequence deviates from the historical measured value by more than a preset range;

[0028] The time sequence of water level changes at upstream and downstream stations caused by the completion sequence violates the direction of water flow propagation.

[0029] Optionally, the uncertainty prediction module includes:

[0030] An ensemble predictor consisting of multiple independently initialized predictive sub-models;

[0031] Training is performed using a weighted objective that combines the Huber loss function with the symmetric mean absolute percentage error loss function;

[0032] During prediction, the statistical distribution of the predicted values ​​is output through the integrated predictor, and the prediction confidence interval is calculated using the Monte Carlo sampling method.

[0033] Optionally, in the trusted evidence storage and verification module, the automated verification rules for the smart contract encoding include:

[0034] Time continuity verification is used to check the consistency of data over time.

[0035] Spatial focus verification is performed by calculating the information entropy of the attention weight distribution output by the spatial attention mechanism in the spatiotemporal feature fusion module to determine the degree of concentration of information fusion.

[0036] Physical compliance verification is used to statistically analyze and determine the overall pass rate of physical constraint verification in a single completion task.

[0037] Optionally, an anomaly propagation blocking mechanism is also included. When the trusted evidence storage and verification module determines that the data of a certain neighboring station is abnormal for multiple consecutive cycles, the spatiotemporal feature fusion module automatically reduces or cuts off the weight of the neighboring station in information fusion and activates a backup associated station.

[0038] Optionally, the uncertainty prediction module also includes an online learning function. When the prediction deviation is detected to continuously exceed the expected range, a lightweight parameter fine-tuning of the prediction model is automatically initiated. The learning rate of the lightweight parameter fine-tuning is lower than the original training learning rate of the model.

[0039] Secondly, this invention provides a method for completing missing hydrological sequence data and predicting trends, comprising the following steps:

[0040] Collect hydrological time-series data from multiple stations within the target watershed and perform edge filtering and quality labeling;

[0041] Local temporal features of the target site are extracted and spatial correlation information of neighboring sites is dynamically fused to form a joint spatiotemporal representation;

[0042] Based on the joint spatiotemporal representation, a preliminary completion sequence for the missing data is generated, and hydrophysical constraints are invoked for iterative verification and correction, and the final completion result is output.

[0043] Using the completed sequence, hydrological elements for future periods are predicted through an integrated prediction model, and the predicted values ​​and corresponding confidence intervals are output.

[0044] The key process data generated in the aforementioned steps are stored on the blockchain for evidence preservation, and automatic compliance verification based on smart contracts is performed.

[0045] It responds to queries and generates a complete data lineage map and decision-making process information that shows the process from raw data to final prediction results.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. By constructing a collaborative modeling network that integrates local temporal features and dynamic spatial correlations, the spatiotemporal collaborative information of multiple stations within the watershed is effectively utilized, significantly improving the robustness and accuracy of data completion in scenarios with high proportions or continuous data loss.

[0048] 2. Innovatively, the data completion process deeply integrates hydrophysical constraints such as the unit hydrograph method and Manning's formula, and designs a correction mechanism based on rule triggering and candidate path reselection to ensure that the completion results not only meet the statistical laws of data, but also strictly conform to the physical consistency of natural hydrological processes, fundamentally avoiding physical distortion.

[0049] 3. A blockchain-based trusted evidence storage and intelligent verification system covering the entire process of data collection, preprocessing, modeling, completion, verification, and prediction has been established. Through consortium blockchains and smart contracts, the system enables tamper-proof recording and automated compliance review of key process data, completely solving the data trust and traceability problems in cross-departmental collaboration.

[0050] 4. It provides full lifecycle visualization and traceability capabilities, which can transparently display key intermediate states such as spatial attention allocation, physical constraint verification details, and model hidden layer activation behind the completion and prediction decisions. This greatly enhances the interpretability of system decisions and the auditability of the process, meeting the high standards of water conservancy management.

[0051] 5. The system integrates engineering-grade robust designs such as anomaly propagation blocking, online incremental learning, and dynamic parameter adjustment, enabling it to adaptively cope with various complex real-world scenarios such as sensor failures, data anomalies, and sudden environmental changes, ensuring the stability and adaptability of the system's long-term operation. Attached Figure Description

[0052] Figure 1 is a schematic diagram of the overall technical solution architecture of the hydrological sequence missing data completion and trend prediction system based on deep learning, blockchain and Internet of Things proposed in this invention.

[0053] Figure 2 is a schematic diagram of the core principle framework of dual-channel spatiotemporal collaborative modeling and physical constraint-driven collaborative completion in this invention;

[0054] Figure 3 is a logical flow diagram of the IoT sensing and edge preprocessing module in this invention.

[0055] Figure 4 is a schematic diagram of the multi-level interaction relationship and data flow of the blockchain trusted evidence storage and intelligent verification module in this invention.

[0056] Figure 5 is a flowchart of the trend prediction module for uncertainty perception in this invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Furthermore, in this invention, an element referred to as fixed to or disposed on another element may be directly disposed on the other element, or there may be an intermediate element. When an element is considered to be connected to another element, it may be directly connected to the other element, or there may be an intermediate element present simultaneously. The terms vertical, horizontal, left, right, and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0059] Example 1

[0060] Referring to Figure 1, the overall system architecture provided in this embodiment of the invention includes six core modules that are connected sequentially and work collaboratively: a data acquisition and preprocessing module, a spatiotemporal feature fusion module, a physical constraint completion module, an uncertainty prediction module, a trusted evidence storage and verification module, and a visualization traceability module. The data flow begins with real-time acquisition from field monitoring stations, is preprocessed at the edge, and then transmitted to the central server. At the central server, spatiotemporal collaborative modeling, physical constraint completion, and uncertainty prediction are performed sequentially. Key data and states throughout the process are stored on the blockchain in real time and verified by smart contracts. Finally, all information can be transparently traced and displayed through a visualization interface. The specific implementation details of each module are as follows.

[0061] This embodiment provides a specific implementation method for the data acquisition and preprocessing module:

[0062] The data acquisition and preprocessing module is the source of the system's data. Its core objective is to complete preliminary quality control and standardization at the edge of data generation. See Figure 3 for the specific implementation process.

[0063] Within the target watershed, monitoring stations such as water level stations, flow stations, and rain gauges are deployed with multi-source heterogeneous sensors, including water level gauges, flow meters, and rain gauges. These sensors synchronously collect raw time-series data at a preset high frequency (e.g., once per minute). The stations are equipped with edge computing units based on low-power microprocessors (e.g., ARM Cortex-M series chips) and LoRa WAN communication modules.

[0064] To suppress random noise introduced by water surface fluctuations or transient disturbances, the edge computing unit performs sliding window averaging filtering on the raw data continuously collected by each sensor. The window length is typically set to 10 sampling points, and the arithmetic mean of the data within the window is taken as the filtered value at that moment.

[0065] The system maintains a sliding window of the most recent 24 hours of historical data for each sensor at the edge unit; it calculates the mean of the data within this window in real time. ) and standard deviation ( When the newly collected raw values satisfy" If a data point is marked as "suspected anomaly", it is then labeled as such. This label serves as a preliminary indicator of data reliability.

[0066] After filtering and anomaly marking, each data record is appended with a structured metadata header, which includes: quality markers (such as normal, suspected anomaly), high-precision timestamp, unique site number, and sensor type code; this generates a preprocessed data stream with a credibility label; this data is then compressed and uploaded to the watershed data center in batches every 15 minutes via the LoRa WAN wide area low power network; this mechanism significantly reduces the network transmission load of invalid or redundant data.

[0067] In one specific implementation, this embodiment provides a specific method for implementing the spatiotemporal feature fusion module:

[0068] This module is deployed on a high-performance computing server in the data center and is responsible for extracting high-order features from the preprocessed data stream for completion and prediction. Its core is a dual-channel parallel network, the structure of which is shown in Figure 2.

[0069] For a target site that needs data completion, the input of this unit is all available time series segments of the site before and after the data missing window that are marked as normal in quality; the input features include water level, flow rate, rainfall, etc., and feature-by-feature Z-score normalization is performed first.

[0070] A three-layer stacked gated recurrent unit (GRU) network is employed. The first GRU layer takes the number of features as input and outputs a hidden state with a dimension of 64, used to capture short-term fluctuations on a minute to hourly scale. The second GRU layer takes the output of the first layer as input and outputs a dimension of 128, used to extract medium-term trends on a timescale of several hours to one day. The third GRU layer further abstracts the data, outputting a final temporal feature vector with a dimension of 256. ； This vector encapsulates all the dynamic characteristics of the target site within the available history window.

[0071] The core of dynamic graph construction is a dynamic graph neural network (Dynamic GNN); the set of nodes in the graph corresponds to all hydrological monitoring stations in the watershed; the edges of the graph represent potential hydrological connections between stations, which are not fully connected, but rather predefined adjacency relationships based on actual river topology (such as upstream and downstream, main stream and tributary connections), but the weights of the edges are dynamic.

[0072] The system automatically updates the graph structure and edge weights of the entire watershed every 6 hours; for any connecting station and sites The edge, its weight Calculated using the following formula that combines statistical correlation and physical hydrodynamic delay:

[0073]

[0074] in: For the site and Within the latest 24-hour sliding window, the standardized Pearson cross-correlation coefficient is calculated based on the same period's water level or flow rate sequence; if a station has missing data within this window, the correlation coefficient from its last valid update period is used for smooth interpolation as a replacement. To from the site Arrive at the station The propagation time of water flow is a physical quantity; the calculation formula is: ,in It is based on the distance between the centerlines of the river channel between two stations pre-stored in a Geographic Information System (GIS). This is the historical average flow velocity of this river section; : Empirical decay constant, used to control the decay rate of the propagation delay effect in the weights, typically with a value of 6 hours.

[0075] Based on the current dynamically weighted graph, a Graph Attention Network (GAT) mechanism is adopted. For the target site, the attention coefficients of all its neighboring sites (i.e., sites connected by edges in the graph) are calculated. The attention coefficients are jointly determined by the edge weights and node features. Then, the latest feature information of all neighboring sites (usually the temporal features extracted by the neighboring sites' own GRU) is weighted and aggregated to generate a spatial embedding vector with a dimension of 256. Joint spatiotemporal representation generation: The 256-dimensional temporal feature vector output by the local feature extraction unit is generated. 256-dimensional spatial embedding vector output by the spatial association fusion unit By concatenating the vectors, a 512-dimensional joint spatiotemporal representation vector is obtained. This vector is a unified, high-order input feature for subsequent completion and prediction tasks.

[0076] In one specific implementation, this embodiment provides a specific implementation method for the physical constraint completion module:

[0077] The physical constraint completion module receives the joint spatiotemporal representation vector. The goal is to generate physically plausible missing data completion values ​​by using available historical data fragments from the target site; its implementation includes sub-units that work collaboratively.

[0078] The sequence generation unit (e.g., sequence-to-sequence decoder) adopts an attention-based encoder-decoder architecture; the encoder encodes available historical data fragments of the target site; the decoder uses a joint spatiotemporal representation vector. As the context for each decoding step, an autoregressive approach is used to gradually generate a preliminary complete sequence for hydrological sequences (such as flow rates) for missing time periods (e.g., 8 consecutive hours, totaling 480 minute points).

[0079] The physical verification unit includes a pre-defined physical constraint model, real-time verification rules, and a collaborative correction mechanism.

[0080] The pre-defined physical constraint model includes: unit hydrodynamic flow constraint and Manning's formula velocity constraint.

[0081] The unit-line flow constraint is expressed in integral form:

[0082] .in The outlet cross-sectional flow rate, This is a clean rain process. For instantaneous unit line, The time delay variable is used to constrain the flow process line to be completed from a macroscopic perspective. This model must conform to the lag, superposition and attenuation characteristics of the watershed rainfall-runoff response.

[0083] Manning's formula for velocity constraints is expressed as ;in The cross-sectional average velocity is... This is the roughness coefficient (the system pre-stores typical value ranges for each river section). The hydraulic radius (which can be estimated from water level and cross-sectional shape) is the radius of gravity. The water surface gradient (which can be estimated from the water level difference and distance between adjacent water level stations); this formula is used to invert and constrain the rate of change of water level. The reasonable upper limit of ", among which Represents unit of time Changes in water level within the area.

[0084] Furthermore, hard-coding or configuring real-time verification rules in the verification unit includes:

[0085] Rule 1: If the absolute value of the water level change rate generated by the decoder at a certain minute exceeds 1.5 times the extreme value of the minute-level change rate under the same hydrological conditions observed in the history of that station, it is judged as a physical anomaly.

[0086] Rule 2: If the confluence time of this flood, calculated by back-calculating the unit hydrograph method based on the preliminary completed flow sequence, deviates from the average confluence time of typical historical floods by more than 2 hours, it is judged as a physical anomaly.

[0087] Rule 3: If the completion of the sequence results in a logical contradiction where the water level rises earlier downstream than upstream between the target station and its direct upstream station, it is determined to be a physical anomaly.

[0088] Furthermore, the collaborative correction mechanism ensures that the physical verification unit and the sequence generation unit operate synchronously, performing real-time verification on the candidate values ​​generated by the sequence generation unit at each step. When any of the above rules is triggered, the module immediately initiates the correction mechanism, which includes the following steps:

[0089] Step 1. Control sequence generation unit backtracks to one of the top-3 optimal alternative paths, different from the current path, reserved by its beam search algorithm.

[0090] Step 2. Regenerate candidate completion values ​​from the alternative path and submit them to the physical verification unit again.

[0091] Step 3. This iterative process continues until the generated complete sequence passes all preset physical consistency rules.

[0092] Step 4. Output the final completion result, forming a complete hydrological sequence that is both statistically and physically continuous and reasonable.

[0093] In one specific implementation, the specific implementation process of the uncertainty prediction module is provided:

[0094] The uncertainty prediction module uses the completed historical sequence to predict future hydrological trends and quantifies the uncertainty of the prediction. Its process framework is shown in Figure 5.

[0095] An ensemble predictor is constructed, consisting of five structurally identical but independently initialized Long Short-Term Memory (LSTM) sub-models. Each sub-model employs a three-layer LSTM layer, with the last layer followed by a Bayesian linear regression output layer.

[0096] The training input is a complete, continuous multi-source hydrological sequence (such as water level, flow rate, and rainfall over the past 72 hours). The output is a prediction of hourly water levels for the next 72 hours.

[0097] Loss function: A weighted loss function is used for end-to-end training.

[0098]

[0099] The loss is insensitive to outliers, improving the model's robustness; To ensure the relative accuracy of the predicted values ​​in terms of magnitude, a symmetrical average absolute percentage error is used.

[0100] In this embodiment, uncertainty is quantified in the following way:

[0101] Bayesian output layer variance: The Bayesian output layer of each LSTM sub-model naturally outputs the mean of the predicted values. and variance .

[0102] Monte Carlo Dropout Sampling: During the prediction inference phase, the Dropout layer in each sub-model is activated, and 50 random forward propagation samples are performed. Due to the randomness of Dropout, each sample will yield slightly different prediction results.

[0103] Integration and Confidence Interval Calculation: First, the predicted means of the five sub-models are averaged to obtain the final overall predicted mean. Then, the distributions generated by 50 Monte Carlo Dropout samplings across all sub-models are combined to calculate the 95% confidence intervals for the water level predictions at future time points. This provides decision-makers with crucial quantitative information on risk.

[0104] This embodiment provides a specific method for online incremental learning functionality:

[0105] When new actual observation data Upon arrival, calculate the deviation between its value and the predicted mean at the corresponding time. If satisfied After the cumulative state reaches 24 hours, the model is determined to have a systematic bias; among which, The predicted value at this moment output by the ensemble predictor The corresponding standard deviation is used to quantify the uncertainty of the forecast.

[0106] Automatically initiate lightweight recalibration. This process fixes the weights of all feature extraction layers (i.e., LSTM layers) in each LSTM sub-model of the ensemble predictor, and only fine-tunes the weights of the Bayesian linear regression output layer of each sub-model.

[0107] The learning rate used for fine-tuning was strictly set to one-fifth (1 / 5) of the original model's training learning rate, and the training data consisted only of newly added observations and completion data from the last 72 hours, in order to balance the model's adaptability to new patterns with its long-term stability.

[0108] In one specific implementation method, the specific implementation of the trusted evidence storage and verification module is described as follows:

[0109] The trusted evidence storage and verification module provides the system with full-process trusted assurance and automated auditing. Its multi-level interaction relationship is shown in Figure 4.

[0110] A consortium blockchain network is built based on the Practical Byzantine Fault Tolerance (PBFT) consensus algorithm. Nodes are jointly deployed and maintained by relevant stakeholders, including river basin management agencies, hydrological bureaus, meteorological bureaus, and environmental protection departments.

[0111] Furthermore, the following is the specific process of putting key process data onto the blockchain:

[0112] The system encapsulates the key process data generated by the aforementioned modules into structured transactions, including but not limited to: the hash value of the original data packet, the feature summary of the joint spatiotemporal representation vector output by the spatiotemporal feature fusion module, the detailed verification flags (which rules are passed / violated) and key parameters of the physical verification unit, and the mean and variance output by the uncertainty prediction module.

[0113] Following the principle of on-chain hashing and off-chain storage, the full text of each transaction (which may be large) is first stored in the decentralized InterPlanetary File System (IPFS), obtaining a unique IPFS Content Identifier (CID). Subsequently, this CID is hashed together with key transaction metadata (such as timestamp, site ID, and transaction type), and this hash is written to the consortium blockchain. This mechanism ensures that the data is immutable and fully traceable.

[0114] The on-chain smart contract encodes three types of automatic verification rules, which are triggered and executed after a transaction is recorded on the chain:

[0115] Check whether the hash values ​​of two data transactions uploaded from the same site and occurring at adjacent times are consecutive on the blockchain to prevent data from being maliciously deleted or inserted.

[0116] Weight distribution of spatial attention mechanism output in the contract-based spatiotemporal feature fusion module Calculate its information entropy ;like If the model's attention is too scattered, it may be due to poor data quality across all related sites, and the transaction will be marked as pending audit.

[0117] Physical compliance verification: In a single completion task, the total number of times the physical verification unit is invoked and the number of times verification passes are statistically analyzed to calculate the pass rate. If the pass rate is below 90%, even if the final result is output, the transaction corresponding to that completion task is marked as pending audit.

[0118] In this embodiment, the following exception handling and blocking mechanisms are provided:

[0119] Any transaction marked as pending audit will automatically trigger an alarm and push the transaction and related context information to a dedicated manual review platform for final review and adjudication by hydrological experts.

[0120] When a smart contract marks a data record from a neighboring site as pending audit for three consecutive standard processing cycles (i.e., 18 hours), the system determines that the data source for that site remains untrustworthy. The spatiotemporal feature fusion module automatically receives this signal and, in its dynamic graph neural network, forces the edge weights between that neighboring site and all target sites to zero. Simultaneously, based on real-time calculations... By ranking the correlation coefficients, the stations with the second-best correlations from the candidate neighbor set are selected to replace their positions in the spatial aggregation, thereby blocking the spread and contamination of low-reliability or abnormal data in the spatial dimension at the source of modeling.

[0121] In one specific implementation, the following describes the specific implementation of the visual traceability module: This module is a human-computer interaction portal for the system and administrators and decision-makers, providing deep process transparency capabilities.

[0122] When a user queries a specific completion and prediction event via the web interface (e.g., querying the data repair and future water level prediction for site S042 on July 15, 2023), the module's backend service immediately starts. It accesses the blockchain explorer and IPFS gateway in parallel via API, aggregating all on-chain transaction hashes, off-chain full-text data, and related information from the database related to the event.

[0123] Using graph database technologies (such as Neo4j), a complete data lineage graph can be automatically constructed. This graph, combining a timeline and a processing flow topology diagram, visually displays every step from raw data acquisition, edge preprocessing, spatiotemporal modeling, physical completion, trend prediction to final evidence storage.

[0124] Users can click on any node or edge in the graph, and the interface will retrieve and display detailed intermediate state information within milliseconds, for example:

[0125] Using the watershed map as the base image, the attention weight of each neighboring station to the target station is overlaid and displayed at the critical moment of completion, with darker colors indicating greater influence.

[0126] The visualization displays the activation value distribution of specific hidden layers of GRU or LSTM at key time steps in the spatiotemporal feature fusion module, aiding in the understanding of the model's internal state.

[0127] List the specific records of the physical verification unit triggering the verification rules, including the violated constraint formula, the calculated value, the set threshold, and whether it ultimately passed.

[0128] Displays the blockchain hash, IPFSCID, smart contract verification result (pass / pending audit) of the corresponding transaction, and the final conclusion of the manual review (such as acceptance or rejection).

[0129] This deep visualization and traceability capability makes the black-box decision-making process of deep learning models, the verification logic of physical constraints, and the blockchain's evidence storage and verification chain completely transparent.

[0130] Example 2

[0131] Application of this invention in routine monitoring of large watersheds

[0132] In a network of 127 stations covering a 50,000 square kilometer basin in the middle and lower reaches of the Yangtze River, the system operated stably according to the specific implementation method described above. Regarding the 8-hour data loss due to a communication interruption at station S042, the system workflow is as follows:

[0133] Data preparation: The data acquisition and preprocessing module uploads the preprocessed data stream.

[0134] Feature Fusion: The spatiotemporal feature fusion module uses available data from 12 hours before and after S042, combined with data based on dynamic formulas. The weights of neighboring stations are calculated (e.g., upstream station A has a weight of 0.85, downstream station B has a weight of 0.45, and lateral tributary station C has a weight of 0.20), generating a joint representation. .

[0135] Physical Completion: The sequence generation unit of the physical constraint completion module generates an initial flow sequence. The physical verification unit finds that a 10-minute flow increase rate violates the Manning formula constraint (Rule 1). The correction mechanism is activated, and the sequence generation unit reverts to the second alternative path and regenerates the sequence, which this time passes all verifications.

[0136] Trend prediction: The uncertainty prediction module uses the completed sequence to output that the water level will rise by 1.2 meters in the next 72 hours, with a 95% confidence interval of [0.9 meters, 1.5 meters].

[0137] Evidence Storage and Verification: The trusted evidence storage and verification module records all data on the blockchain. The smart contract calculates the spatial attention entropy to be 0.45 (<0.7), and the physical verification pass rate is 100% (>90%), thus the verification is successful.

[0138] Traceability Display: Through the visual traceability module interface, the administrator can clearly see that the completion of S042 mainly depends on the upstream station A (attention weight 0.62), and the physical verification is complete, indicating high decision credibility.

[0139] Example 3

[0140] Application of this invention in typhoon emergency response in coastal cities

[0141] In the flood control system of a city in southeastern coastal China, when a typhoon struck, data from multiple river monitoring points was frequently lost due to communication interference caused by the heavy rain.

[0142] Parameters are dynamically adjusted:

[0143] Data acquisition and preprocessing module: Automatically switches to emergency mode, increases the acquisition frequency to once every 30 seconds, adjusts the filter window to 5 points, and dynamically calculates the 3σ criterion based on data from the past 2 hours.

[0144] Spatiotemporal feature fusion module: The graph network update frequency is increased to once per hour, and the decay constant is... The interval was adjusted to 3 hours, making the system focus more on recent, rapid correlations.

[0145] Physical constraint completion module: Dynamically adjusts the roughness coefficient in the Manning formula based on the rainfall intensity retrieved from radar. (The value of n decreases during heavy rain); the unit line rule uses the instantaneous unit line of a historically similar typhoon, Hagupit, as a benchmark.

[0146] Addressing high-concurrency loss: When multiple adjacent sites lose connection simultaneously, the spatiotemporal feature fusion module sets them as hidden nodes in the graph network. By utilizing information from upstream and downstream sites that are not lost connection, the module estimates their state through the message passing mechanism of the graph neural network, maintaining the connectivity of the spatial connection network and achieving blind replacement in extreme cases.

[0147] Online learning adaptation: After the typhoon made landfall, the actual water level rose faster than the model expected. Eighteen deviation alarms were triggered within 24 hours (>15 thresholds). The online incremental learning function of the uncertainty prediction module was activated, fine-tuning the output layers of the five LSTM sub-models. This enabled the model to quickly adapt to the extreme hydrological response patterns brought by the typhoon, significantly improving subsequent prediction accuracy.

[0148] Emergency Auditing and Decision-Making: The trusted evidence storage and verification module marks all operations and decisions during emergencies with emergency modes and records them on the blockchain. The physical compliance verification threshold for smart contracts is temporarily relaxed to 85% to accommodate decisions under higher uncertainty. At the flood control command center, decision-makers can use the visualization and traceability module to instantly access the overlay map of predicted confidence intervals for flood-prone areas throughout the city, clearly seeing which areas have high prediction uncertainty and require strengthened defenses, thus achieving data-driven, risk-visible, and process-traceable scientific emergency command.

[0149] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0150] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A system for completing missing hydrological series data and predicting trends, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source hydrological time-series data and perform edge-side preprocessing. The spatiotemporal feature fusion module is used to extract the temporal features of the target station and fuse the spatial information of neighboring stations to generate a joint spatiotemporal representation; the physical constraint completion module is used to generate a completed sequence based on the joint spatiotemporal representation and to verify and correct it using hydrological physical constraints; the uncertainty prediction module is used to perform integrated prediction based on the completed sequence and output the prediction result and its confidence interval. The trusted evidence storage and verification module is used to store key process data of the system on the blockchain and perform automated verification through smart contract coding. It also includes a visualization and traceability module, which provides full lifecycle data lineage tracing and visualization from data collection to predictive output.

2. The system according to claim 1, characterized in that, The spatiotemporal feature fusion module includes: a local feature extraction unit, which uses a recurrent neural network to extract the temporal dependency features of the target station; and a spatial association fusion unit, which models the hydrological association between stations based on the structure of a dynamic graph and aggregates neighboring station information through an attention mechanism; wherein, the edge weights in the dynamic graph are determined based on the statistical correlation between the data of the stations and the physical delay of water flow propagation.

3. The system according to claim 2, characterized in that, In the spatial association fusion unit, the edge weights Calculated dynamically using the following formula: Where t represents the center time of the current processing time window, Indicates site and Pearson correlation coefficient within the sliding time window This represents the estimated water propagation time based on the river channel distance and average flow velocity between two stations. This is the empirical decay constant.

4. The system according to claim 3, characterized in that, The physical constraint completion module includes: a sequence generation unit for generating a preliminary completion sequence based on the joint spatiotemporal representation; a physical verification unit, which integrates unit line flow constraint and / or Manning formula velocity constraint, for performing real-time rationality verification on the preliminary completion sequence; and a correction unit for triggering the sequence generation unit to regenerate a candidate completion sequence when the physical verification unit determines an anomaly, until the output completion sequence meets the preset physical verification rules.

5. The system according to claim 4, characterized in that, The physical verification unit has preset verification rules including at least one of the following: the rate of change of water level in the completed sequence exceeds the threshold determined based on historical data; the basin confluence time inferred from the completed sequence deviates from the historical measured value by more than a preset range; and the time sequence relationship of water level changes at upstream and downstream stations caused by the completed sequence violates the direction of water flow propagation.

6. The system according to claim 5, characterized in that, The uncertainty prediction module includes: an ensemble predictor consisting of multiple independently initialized prediction sub-models; trained using a weighted objective combining the Huber loss function and the symmetric mean absolute percentage error loss function; during prediction, the ensemble predictor outputs the statistical distribution of the predicted values ​​and calculates the prediction confidence interval using the Monte Carlo sampling method.

7. The system according to claim 6, characterized in that, In the trusted evidence storage and verification module, the automated verification rules for the smart contract encoding include: time continuity verification, used to check the continuity of data in the time series; spatial focus verification, which judges the degree of concentration of information fusion by calculating the information entropy of the attention weight distribution output by the spatial attention mechanism in the spatiotemporal feature fusion module; and physical compliance verification, used to statistically determine the overall pass rate of physical constraint verification in a single completion task.

8. The system according to claim 7, characterized in that, It also includes an abnormal propagation blocking mechanism. When the trusted evidence storage and verification module determines that the data of a certain neighboring station is abnormal for several consecutive cycles, the spatiotemporal feature fusion module automatically reduces or cuts off the weight of the neighboring station in information fusion and activates a backup associated station.

9. The system according to claim 8, characterized in that, The uncertainty prediction module also includes an online learning function. When the prediction deviation is detected to be continuously exceeding the expected range, it automatically starts a lightweight parameter fine-tuning of the prediction model. The learning rate of the lightweight parameter fine-tuning is lower than the original training learning rate of the model.

10. A method for completing missing hydrological series data and predicting trends, characterized in that, Includes the following steps: Collect hydrological time-series data from multiple stations within the target watershed and perform edge filtering and quality labeling; Local temporal features of the target station are extracted and spatial correlation information of neighboring stations is dynamically fused to form a joint spatiotemporal representation. Based on the joint spatiotemporal representation, a preliminary complete sequence for missing data is generated, and hydrophysical constraints are invoked for iterative verification and correction, and the final complete result is output. Using the complete sequence, hydrological elements for future periods are predicted through an integrated prediction model, and the predicted values ​​and corresponding confidence intervals are output. The key process data generated in the aforementioned steps are stored on the blockchain for evidence preservation, and automatic compliance verification based on smart contracts is performed. It responds to queries and generates a complete data lineage map and decision-making process information that shows the process from raw data to final prediction results.

Images