A river and lake ecological water supplement prediction method based on intelligent sensing
By using intelligent sensing technology and multi-layer models in river and lake ecological water replenishment, the uncertainty and supply-demand mismatch in ecological water demand prediction of existing methods have been solved, achieving accurate ecological water demand prediction and water replenishment scheme optimization, thereby improving water resource utilization efficiency and ecological protection effectiveness.
Patent Information
- Application Number
- CN202511486501.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-04-28
- Estimated Expiration
- 2045-10-17
AI Technical Summary
Existing methods for ecological water replenishment of rivers and lakes fail to effectively consider the specific needs and uncertainties of biological life cycles, resulting in ecologically invalid prediction results or failed schemes. Furthermore, the supply and demand are mismatched, reducing water resource utilization efficiency and ecological protection benefits.
A smart sensing-based approach is adopted, which involves setting up data acquisition sensors in the watershed to label ecological mechanisms and uncertainties, establishing a multi-dimensional data linkage mechanism, using Bayesian networks to update uncertainty weights, combining Bayesian networks, LSTM and XGBoost models to predict ecological water demand, and optimizing water replenishment schemes through a digital twin decision system.
It achieves both accuracy and flexibility in ecological water demand forecasting, enabling it to cope with uncertainties, ensure the achievement of ecological goals, and improve water resource utilization efficiency and ecological protection benefits.
Smart Images

Figure CN121390420B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of river and lake ecological water replenishment technology, specifically a river and lake ecological water replenishment prediction method based on intelligent sensing. Background Technology
[0002] Ecological water replenishment for rivers and lakes is an important aspect of water resource allocation and ecological protection, playing a crucial role in maintaining the watershed's aquatic ecological balance and restoring damaged aquatic environments. Currently, the industry has largely completed the construction of basic water replenishment facilities, and the technological focus is gradually shifting towards refined management. However, the following technical challenges still exist:
[0003] Existing methods generally build models based on pure hydrological parameters such as water level and flow, simplifying ecological water demand to a fixed threshold. This ignores the specific needs of biological life cycles, such as the need for a water flow pulsation frequency of no less than 0.5 times / hour during the fish spawning period and the need for water temperature and dissolved oxygen to meet standards during the aquatic plant regrowth period. It also does not consider the interaction effects between hydrology, soil and vegetation. For example, when the soil moisture content is less than 12% during drought, vegetation transpiration will exacerbate the consumption of river and lake water, requiring an additional 10% to 15% of water to be added. Ultimately, this results in the predicted values meeting the standards but the ecology being ineffective.
[0004] Extreme weather and sudden changes in hydrological conditions within the basin, as well as human uncertainties such as temporary industrial water intake and emergency water diversion for agricultural irrigation, can all cause sudden changes in water demand or supply capacity. However, existing models do not transform this uncertain information into prediction boundaries and only output a single deterministic result. Once an uncertain event occurs, the plan is prone to failure and lacks risk warning.
[0005] Existing technologies either focus on local water demand forecasting or only pay attention to the allocation of total water supply at the basin level, without linking the water supply capacity of sub-basins with the ecological priority of river sections. This results in problems such as local water demand exceeding the standard but global water supply being sufficient, or global water supply being insufficient but local water replenishment being excessive, which reduces the efficiency of water resource utilization and the benefits of ecological protection. Summary of the Invention
[0006] The purpose of this invention is to provide a method for predicting ecological water replenishment of rivers and lakes based on intelligent sensing, so as to solve the problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting ecological water replenishment of rivers and lakes based on intelligent sensing, the specific steps of which are as follows:
[0008] Step S1: Set up data acquisition sensors in the watershed and label the sensors with ecological mechanism tags and uncertainty tags. Adjust the sampling frequency according to the uncertainty level, establish a multi-dimensional data linkage mechanism for biological anomaly triggering, and output the raw data.
[0009] Step S2: Update the uncertainty weights using a Bayesian network, verify the validity of hydrological parameters using biological response data, remove outliers and fill in missing values, and standardize the units using Z-Score.
[0010] Step S3: First calculate the importance of features, then select the core features, establish a three-level feature linkage matrix of river section, sub-basin, and basin, divide the feature time period according to the biological life cycle stage, align the hydrological and ecological data and label the associated tags;
[0011] Step S4: Establish a three-layer linkage model. The mechanism constraint layer incorporates a biological stress response sub-model. The data-driven prediction layer uses a combination model of LSTM and XGBoost to output the baseline ecological water demand. The uncertainty quantification layer introduces variational inference to adjust the water demand fluctuation range.
[0012] Step S5: Based on the model in Step S4, perform model training and transfer optimization, automatically perform incremental updates according to ecological and risk thresholds, and train the model offline so that the model can learn data patterns and ecological mechanisms simultaneously.
[0013] Step S6: Prioritize water supply by combining ecological importance with uncertainty risk formula, establish multi-objective robust optimization function, use improved NSGA-III and reinforcement learning to solve 3-5 optimal solutions, and allocate water supply according to watershed, sub-watershed and river segment to form a multi-scale water replenishment scheme.
[0014] Step S7: Establish a digital twin decision-making system to simulate the ecological response after water replenishment, set up an automatic biological data feedback module, and design an interface linking uncertainty and ecological impact.
[0015] Furthermore, step S1 is specifically as follows:
[0016] Three monitoring sections were set up within the watershed. Each section was equipped with a sensing terminal and simultaneously labeled with ecological mechanism tags and uncertainty tags: underwater acoustic sensors labeled fish spawning period and upstream water fluctuations; soil moisture sensors labeled vegetation transpiration and agricultural emergency water diversion; water level and flow sensors labeled hydrological conditions and extreme weather; dissolved oxygen sensors labeled dissolved oxygen threshold for fish spawning and algal bloom interference; underwater cameras labeled fish spawning volume, egg and larval survival status and water turbidity interference; and vegetation spectral sensors labeled vegetation chlorophyll content, coverage and photosynthetic efficiency, realizing the binding of mechanism and risk association during the data collection stage.
[0017] The sampling frequency is adjusted according to the uncertainty level. When meteorological satellites detect extreme weather warnings or when government interfaces receive strong interference from industrial water intake, the sampling frequency of sensors in the corresponding areas is increased to 5 minutes / time, while the sampling frequency in non-interference areas is maintained at 1 hour / time. Data transmission uses EC-500 edge computing nodes in combination with 5G slicing.
[0018] When the underwater camera detects a sudden drop in fish spawning, a chlorophyll content below 35 mg / m² as monitored by the vegetation spectral sensor, a sudden drop in fish activity frequency, or abnormal vegetation chlorophyll content, it automatically triggers synchronous high-frequency data collection from related sensors, including upstream flow, dissolved oxygen, and surrounding water intake gates, forming a complete data chain of biological anomalies, hydrological parameters, and interference sources.
[0019] Furthermore, step S2 is specifically as follows:
[0020] Based on the multi-source data from step S1, the data is processed according to the dual requirements of ecological mechanism and uncertainty. A Bayesian network is used to update the uncertainty weight in real time. The initial labeling is weak interference from agricultural water diversion. If the water diversion volume is detected to exceed the registered value by 50%, the weight is updated to 0.6, and the associated vegetation transpiration water demand data is marked as high uncertainty associated data. The validity of hydrological parameters is verified using biological response data. If the hydrological data shows a water flow pulsation frequency of 0.6 times / hour, but the underwater acoustic sensor deployed in step S1 detects that the fish spawning rate is less than 30%, it is determined that there is a mechanism deviation, triggering a second sampling and correction of the dissolved oxygen sensor.
[0021] By constraining ecological thresholds and removing outliers using isolated forests, and by filling in missing values through mechanistic similarity interpolation, a dual-association dataset is finally formed through Z-Score standardization and unified dimensionality.
[0022] Furthermore, step S3 is specifically as follows:
[0023] Based on the dual-association dataset from step S2, features are selected using a dual criterion of SHAP value and ecological mechanism. First, the importance of features is calculated through random forest, and then the influence of features on the uncertainty interval is analyzed through SHAP value. Irrelevant features are eliminated, and finally 15-20 core features are retained.
[0024] A three-level feature linkage matrix was established for river sections, sub-basins, and watersheds. When the water demand for fish spawning at the river section level is correlated with the water storage capacity of reservoirs at the sub-basin level, an uncertainty transmission coefficient was introduced. When the sub-basin is in drought with high uncertainty, the coefficient was set to 1.2, which means the water demand was increased by 20%. At the same time, feature time periods were divided according to the biological life cycle stage. The Pandas library of Python was used to slice and align the hydrological and ecological data according to the key time periods and label the correlation.
[0025] Furthermore, step S4 is specifically as follows:
[0026] Based on the core features selected in step S3, a three-layer linkage model is established. The mechanism constraint layer incorporates the biological stress response sub-model. On the basis of the basic ecological water demand model during the fish spawning period, stress-corrected water demand is added to form a two-dimensional model of basic water demand and stress water demand. Results that do not conform to the mechanism are forcibly corrected.
[0027] The data-driven layer adopts a combined model of LSTM and XGBoost. The LSTM is set to a 3-layer structure to capture the temporal data dependencies including water level and evapotranspiration. The XGBoost is set to a learning rate of 0.1 and a tree depth of 8 to fit the nonlinear relationship between soil moisture content and water demand. The core features of step S3 are input to output the baseline ecological water demand.
[0028] The uncertainty quantification layer introduces variational inference, dynamically adjusting the water demand fluctuation range hourly based on biological anomaly data and uncertainty data, and finally outputting the baseline water demand, confidence interval, and ecological interpretation label.
[0029] Furthermore, step S5 is specifically as follows:
[0030] Based on the three-layer linkage model in step S4, the core mechanism parameters are fixed during cross-scale migration, and the stress coefficient of fish and vegetation transpiration coefficient at the river section level are used as anchor parameters. When migrating to the sub-basin, only the uncertainty transmission coefficient of the data-driven layer is finely adjusted to ensure the consistency of biological mechanisms in the same basin.
[0031] When the monitoring shows that the wilting rate of aquatic plants exceeds 20% or the mean uncertainty weight exceeds 0.6, the trigger delay is set to 15 minutes, and only the uncertainty quantification layer and the data-driven layer are updated, while the parameters of the mechanism constraint layer are retained.
[0032] The offline training phase uses full data from the past 10 years, including hydrological, ecological, and uncertain data, and runs on an RTX A6000 GPU server. The dataset is divided into layers according to life cycle stage and uncertainty level. A comprehensive loss function with penalty terms for prediction bias and mechanism bias is designed, and the loss is minimized by the AdamW optimizer to ensure that the model learns data patterns and ecological mechanisms simultaneously.
[0033] Furthermore, step S6 is specifically as follows:
[0034] Based on the baseline water demand output by the model trained in step S5, supply and demand are optimized in a coordinated manner. A formula for prioritizing ecological importance and uncertainty risks is established to achieve dynamic ranking. A multi-objective robust optimization function is established with the objectives of minimizing the water demand gap in sensitive areas, minimizing water waste, and maximizing the robustness of the scheme. Three levels of constraints are set: water demand in sensitive areas at the river section level ≥ baseline value minus uncertainty lower limit; water supply at the sub-basin level ≥ total water demand in sensitive areas within the jurisdiction; total water supply at the basin level ≥ total water demand in all sub-basins, and the gap ≤ 3% under the scenario of extreme drought and strong disturbance.
[0035] An improved NSGA-III algorithm combined with reinforcement learning is used to solve the problem, outputting a set of 3-5 Pareto optimal solutions. The schemes are allocated at the basin, sub-basin, and river section levels. At the basin level, the water supply quota for the sub-basin is determined by referring to the ecological water replenishment effect data of the same period in the past 3 years. At the sub-basin level, the quota is allocated to the river section to form a multi-scale water replenishment scheme.
[0036] Furthermore, step S7 is specifically as follows:
[0037] Based on the scaled water replenishment plan in step S6, a digital twin decision-making system is established. A three-dimensional visualization model is constructed using Unity3D and combined with GIS geographic data. The water replenishment plan is input into the twin model to simulate the ecological response, including fish spawning rate and vegetation survival rate after water replenishment. If the simulated vegetation survival rate is less than 60%, the system automatically returns to step S6 for re-optimization, forming a virtual-real closed loop.
[0038] Deploy an automatic biological data feedback module to collect biological data in real time after water replenishment using the underwater camera and vegetation spectral sensor deployed in step S1. The data is uploaded to the cloud via the MQTT protocol. The ecological effect deviation rate is automatically calculated weekly, and the deviation value is input into the model training stage in step S5.
[0039] The interface links uncertainty and ecological impact. The left side displays the biological stress status at the river section level, the middle shows the distribution of uncertainty sources at the sub-basin level, and the right side presents suggestions for adjusting the watershed-level water replenishment plan. It also has the function of clicking to view the correlation chain between mechanism, uncertainty, and ecological effect.
[0040] The beneficial effects of this invention are as follows:
[0041] 1. This invention involves deploying sensors within the watershed that are labeled with ecological mechanisms. Underwater acoustic sensors are linked to fish spawning periods, and soil moisture sensors are linked to vegetation transpiration, establishing a direct link between organisms and hydrology during the data acquisition phase. Subsequently, characteristic time periods are divided according to the biological life cycle, aligning hydrological and ecological data. A two-dimensional model including basic water demand and stress water demand is also established, and the validity of hydrological parameters is verified by combining biological response data. For example, when fish spawning rates are detected to be too low, relevant parameters are corrected. This ensures that ecological water demand predictions fully match the actual needs of organisms at different stages, ensuring that ecological goals such as fish spawning rates and vegetation survival rates are truly achieved after water replenishment.
[0042] 2. This invention addresses uncertainty throughout the entire process from data acquisition to solution output. It adjusts sensor sampling frequencies based on uncertainty levels, significantly increasing sampling frequencies in areas experiencing extreme weather or strong industrial interference with water intake to proactively capture risk signals. It also uses a Bayesian network to update uncertainty weights in real time, marking high-risk associated data. Furthermore, it incorporates variational inference technology to adjust water demand fluctuation ranges hourly based on biological anomaly data and uncertainty data. Uncertainty risk considerations are integrated into priority ranking and solution optimization, transforming uncertainty into quantifiable predictable boundaries and outputting results with confidence intervals to provide early warnings of risks and prevent solution failure due to unforeseen circumstances.
[0043] 3. This invention establishes a three-tiered characteristic linkage matrix at the river segment, sub-basin, and watershed levels, using an uncertainty transmission coefficient to connect local demand with global supply. In water replenishment allocation, the goal is to minimize water shortages in sensitive areas, minimize water waste, and maximize the robustness of the scheme. Allocation is made hierarchically, with watersheds determining sub-basin quotas based on historical replenishment effects. When allocating from sub-basins to river segments, buffer zones are appropriately increased in sensitive areas, while ordinary river segments receive a reasonable reduction. Furthermore, a digital twin system simulates the ecological response after water replenishment; if the simulation is unsatisfactory, it automatically returns to optimization, achieving precise matching between local demand and global supply. This avoids water waste while ensuring ecological needs at all levels, simultaneously improving water resource utilization efficiency and ecological protection benefits. Attached Figure Description
[0044] Figure 1 This is a flowchart of the river and lake ecological water replenishment prediction method based on intelligent sensing according to the present invention.
[0045] Figure 2 This is a flowchart illustrating the construction and training process of the three-layer linkage model of this invention.
[0046] Figure 3 This is a flowchart illustrating the optimization and feedback process of the scaled water replenishment scheme of the present invention. Detailed Implementation
[0047] 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.
[0048] like Figures 1 to 3 As shown in the figure, this embodiment of the invention provides a method for predicting ecological water replenishment of rivers and lakes based on intelligent sensing. The specific steps of the method are as follows:
[0049] Step S1: Set up data acquisition sensors in the watershed and label the sensors with ecological mechanism tags and uncertainty tags. Adjust the sampling frequency according to the uncertainty level, establish a multi-dimensional data linkage mechanism for biological anomaly triggering, and output the raw data.
[0050] Step S2: Update the uncertainty weights using a Bayesian network, verify the validity of hydrological parameters using biological response data, remove outliers and fill in missing values, and standardize the units using Z-Score.
[0051] Step S3: First calculate the importance of features, then select the core features, establish a three-level feature linkage matrix of river section, sub-basin, and basin, divide the feature time period according to the biological life cycle stage, align the hydrological and ecological data and label the associated tags;
[0052] Step S4: Establish a three-layer linkage model. The mechanism constraint layer incorporates a biological stress response sub-model. The data-driven prediction layer uses a combination model of LSTM and XGBoost to output the baseline ecological water demand. The uncertainty quantification layer introduces variational inference to adjust the water demand fluctuation range.
[0053] Step S5: Based on the model in Step S4, perform model training and transfer optimization, automatically perform incremental updates according to ecological and risk thresholds, and train the model offline so that the model can learn data patterns and ecological mechanisms simultaneously.
[0054] Step S6: Prioritize the water supply by combining ecological importance with the uncertainty risk formula, establish a multi-objective robust optimization function, and use the improved NSGA-III and reinforcement learning to solve 3-5 optimal solutions. Allocate the water supply according to the watershed, sub-watershed, and river segment to form a multi-scale water replenishment scheme.
[0055] Step S7: Establish a digital twin decision-making system to simulate the ecological response after water replenishment, set up an automatic biological data feedback module, and design an interface linking uncertainty and ecological impact.
[0056] Specifically, step S1 is as follows:
[0057] Three monitoring sections were set up within the watershed, located at key nodes or ecologically sensitive areas in the upper, middle and lower reaches of the watershed, such as fish spawning grounds and core vegetation areas, to ensure coverage of the main hydrological and ecological units of the watershed. Each section was equipped with a sensing terminal and simultaneously labeled with ecological mechanism tags and uncertainty tags. The underwater acoustic sensor was selected as HY-AS200, labeled with the fish spawning period and upstream water fluctuations. The soil moisture sensor was selected as TDR-300, labeled with vegetation transpiration and agricultural emergency water diversion. The water level and flow sensor was labeled with hydrological conditions and extreme weather. The dissolved oxygen sensor was selected as DO-200, labeled with the dissolved oxygen threshold (≥8mg / L) for fish spawning and algal bloom interference, to realize the mechanism and risk correlation binding in the data acquisition stage.
[0058] The underwater camera used was the UC-4K model, and the data were labeled with fish spawning volume, egg and larval survival status, and water turbidity interference; uncertainty label: water turbidity caused by heavy rain obstructing the monitoring field of view;
[0059] The VS-1000 vegetation spectral sensor was selected, and the vegetation chlorophyll content, coverage and photosynthetic efficiency were labeled; uncertainty labels: water shortage stress of vegetation after agricultural water diversion and industrial dust cover affect the accuracy of spectral monitoring.
[0060] The sampling frequency is adjusted according to the uncertainty level. When meteorological satellites detect extreme weather warnings (high uncertainty, weight ≥ 0.8) or the government interface receives strong interference from industrial water intake (≥ 100,000 cubic meters per day), the sampling frequency of the corresponding sensor is increased to 5 minutes / time, while the sampling frequency in non-interference areas is maintained at 1 hour / time. Data transmission adopts EC-500 edge computing nodes with 5G slicing, and independent 5G slices are selected to ensure the real-time transmission and data security of core hydrological and ecological data, and to ensure the real-time performance of core data.
[0061] Simultaneously, a multi-dimensional data linkage mechanism is established to trigger biological anomalies. When the underwater camera captures a sharp drop in fish spawning (less than 50% of the historical average for the same period in a single day), the chlorophyll content monitored by the vegetation spectral sensor is lower than 35 mg / m² (vegetation stress threshold), the fish activity frequency drops sharply, or the vegetation chlorophyll content is abnormal, the system automatically triggers high-frequency data collection from related sensors, including upstream flow, dissolved oxygen, and surrounding water intake gates. High-frequency data collection continues until the abnormal state is alleviated. If the abnormality persists, the data collection needs are reassessed after 2 hours of fixed data collection, forming a complete data chain of biological anomalies, hydrological parameters, and interference sources.
[0062] Specifically, step S2 is as follows:
[0063] Based on the multi-source data from step S1, including fish activity data from the HY-AS200 sensor and soil moisture data from the TDR-300 sensor, the data was processed according to the dual requirements of ecological mechanism and uncertainty. A Bayesian network was used to update the uncertainty weight in real time, with an iteration cycle of 30 minutes, which was adapted to the sampling frequency of 5 minutes / 1 hour in step one. Initially, the data was labeled as weak interference from agricultural water diversion with a weight of 0.2. If the water diversion volume was detected to exceed the registered value by 50%, the weight was updated to 0.6, which is considered strong interference. The associated vegetation transpiration water demand data was then marked as high uncertainty associated data. The validity of hydrological parameters was verified using biological response data. If the hydrological data showed a water flow pulsation frequency of 0.6 times / hour, but the underwater acoustic sensor deployed in step S1 detected a fish spawning rate of less than 30%, a mechanism deviation was determined, triggering a second sampling and correction by the dissolved oxygen sensor. The second sampling lasted for 2 hours to ensure that the data values met the standards and were ecologically valid.
[0064] By constraining ecological thresholds and removing outliers from isolated forests (e.g., water flow pulsations below 0.2 times / hour during fish spawning season), missing values are filled in by interpolation based on mechanism similarity. Data on soil moisture content of the same vegetation type and similar data from the same period are referenced. Finally, the Z-Score is used to standardize the data and unify the dimensions. The standardization benchmark uses monitoring data from the same period of the past three years for the entire watershed to ensure the rationality of the unified dimensions of data from different types of sensors, thus forming a dual-association dataset.
[0065] Specifically, step S3 is as follows:
[0066] Based on the dual-association dataset from step S2, features are selected using a dual criterion of SHAP value and ecological mechanism. First, the importance of features is calculated using random forest, for example, the frequency score of water flow pulsation during the spawning period is 0.85. Then, the influence of features on the uncertainty interval is analyzed using SHAP value (≥0.7 is retained; this threshold is determined based on effective cases of ecological water demand prediction in the past 5 years to ensure the contribution of retained features to the prediction results). Irrelevant features such as wind speed are removed, and finally 15-20 core features are retained.
[0067] The random forest has 200 decision trees and a maximum depth of 10 layers, with parameter settings that balance computational efficiency and accuracy.
[0068] A three-level feature linkage matrix was established at the river segment, sub-basin, and basin levels. When the water demand for fish spawning at the river segment level is correlated with the water storage capacity of reservoirs at the sub-basin level, an uncertainty transmission coefficient was introduced. When the sub-basin is in a state of high uncertainty due to drought, a coefficient of 1.2 was adopted (the criteria for high uncertainty due to drought are soil moisture content ≤12% or monthly precipitation anomaly percentage ≤-40%), which means the water demand is increased by 20%, thus solving the problem of supply and demand coordination gap. At the same time, feature periods were divided according to the biological life cycle stages, including the fish spawning period from March to April, the fattening period from July to August, the vegetation greening period from April to May, and the withering period from October to November. The Pandas library of Python was used to slice and align hydrological and ecological data according to key periods (such as the spawning period from 6 to 8 am), and label them with association tags such as water flow pulsation frequency of 0.5-1 times / hour and dissolved oxygen ≥8 mg / L, so that the features carry ecological significance.
[0069] Specifically, step S4 is as follows:
[0070] Based on the core features selected in step S3, such as the frequency of water flow pulsation during the spawning period, a three-layer linkage model is established. The mechanism constraint layer incorporates the biological stress response sub-model. On the basis of the basic ecological water demand model during the fish spawning period, stress-corrected water demand is added to form a two-dimensional model of basic water demand and stress water demand. Results that do not conform to the mechanism are forcibly corrected.
[0071] The data-driven layer employs a combined LSTM and XGBoost model. The LSTM is configured with a 3-layer structure: 16 neurons in the input layer, 32 neurons in the hidden layer, and 1 neuron in the output layer. The number of neurons was optimized using 5-fold cross-validation to balance prediction accuracy and computational efficiency, avoiding overfitting or underfitting. It captures the dependence of time-series data such as water level and evapotranspiration. The XGBoost is configured with a learning rate of 0.1 and a tree depth of 8 (parameters optimized using cross-validation) to fit the nonlinear relationship between soil moisture content and water demand. The core features from step S3 are input to output the baseline ecological water demand.
[0072] The uncertainty quantification layer introduces variational inference, dynamically adjusting the water demand fluctuation range hourly based on biological anomaly data and uncertainty data. The fluctuation range is calculated based on a 90% confidence interval using the formula [baseline water demand - 1.64 × data standard deviation, baseline water demand + 1.64 × data standard deviation]. For example, the original 10±20,000 cubic meters was updated to 12±30,000 cubic meters after strong interference water intake was detected, with a confidence level of 90%. The final output includes the baseline water demand, confidence interval, and ecological interpretation labels. For example, to match the water flow pulsation demand during fish spawning season, the labels are divided into four categories: water flow pulsation adaptability, dissolved oxygen compliance probability, water temperature suitability, and water turbidity compliance. The probability of dissolved oxygen concentration meeting the standard is 92%, providing complete prediction results for subsequent model training.
[0073] Formula for calculating basic ecological water requirements during fish spawning season:
[0074]
[0075]
[0076] In the formula: This indicates the basic ecological water requirement during the fish spawning period, in units of 10,000 cubic meters per day. It refers to the amount of water replenishment required to meet the basic hydrological conditions for fish spawning.
[0077] The influence coefficient of water flow pulsation frequency is calibrated using successful fish spawning cases collected in the past 5 years by the HY-AS200 underwater acoustic sensor deployed in step S1. The value range is usually 0.3~0.8, reflecting the contribution weight of water flow pulsation to the water requirement for spawning.
[0078] This represents the frequency function of water flow pulsations, in units of times per hour. This refers to the actual frequency of water flow pulsations monitored, such as during fish spawning season. 0.5 times / hour Through the Linear fitting was performed to obtain the quantitative impact of key hydrological conditions during the spawning period.
[0079] The dissolved oxygen influence coefficient is also calibrated based on successful spawning cases in the past 5 years, with a value range of 0.1-0.4, reflecting the regulatory effect of dissolved oxygen concentration on water demand;
[0080] This indicates the dissolved oxygen concentration in the water, in mg / L. It is monitored in real-time by dissolved oxygen sensors deployed in step S1. During the fish spawning season... 8 mg / L;
[0081] This indicates the slope's effect on the frequency of water flow pulsations, reflecting the effect of each increase in the frequency of water flow pulsations by one per hour. The increment was calibrated by linear regression of the correlation data between the frequency of water flow pulsations during the spawning season and the actual water demand collected in step S1 over the past 5 years, with a value range of 0.4 to 0.6, such as the spawning season of cyprinid fish. It is usually taken as 0.5, which means Each increase per hour The corresponding increase of 0.5 will lead to a corresponding increase in water demand. 0.5;
[0082] This represents the intercept term, used to correct for water demand deviations in the low-frequency range and avoid prediction distortion caused by extreme values. It is also calibrated based on historical data from the past 5 years, with a value range of 0.1 to 0.2. =0.15, to ensure When the frequency is 0.5 times / hour, It can still output effective baseline values, ensuring the continuity of water demand forecasting.
[0083] Formula for calculating water requirement adjusted for stress during fish spawning season:
[0084]
[0085] In the formula: This indicates the water requirement for stress correction during the spawning season of fish, in units of 10,000 cubic meters per day. It refers to the additional water required to offset the adverse effects of stress factors on spawning when the aquatic environment becomes abnormal (such as a sudden drop in water quality or increased disturbance).
[0086] This represents the stress correction baseline coefficient, in units of 10,000 cubic meters per day. It was calibrated using nearly 5 years of fish stress experiment data, and its value is related to the fish species in the watershed, such as cyprinids. The daily volume is typically 20,000 to 50,000 cubic meters, reflecting the fish's sensitivity to environmental stress.
[0087] This represents the fish stress coefficient, ranging from 0 to 1. It is calculated using biological monitoring data from step S1 (such as fish activity frequency and egg / larval survival rate). It is used when water quality suddenly drops. It can rise to 0.3~0.5 without stress. =0, quantifying the impact of environmental stress on water demand.
[0088] Specifically, step S5 is as follows:
[0089] Based on the three-layer linkage model in step S4, the core mechanism parameters are fixed during cross-scale migration. The stress coefficient of fish and vegetation transpiration coefficient at the river section level are used as anchor parameters. When migrating to the sub-basin / basin model, only the uncertainty transmission coefficient of the data-driven layer is fine-tuned. The fine-tuning range is determined according to the hydrological similarity between the sub-basin and the river section: ±5% when the similarity is ≥80% and ±10% when the similarity is <80%, to avoid mechanism distortion and ensure the consistency of biological mechanisms in the same watershed, which is consistent with the watershed uniformity of the perception label in step S1.
[0090] Incremental updates are triggered based on ecological and risk thresholds. When the monitoring shows that the aquatic plant withering rate exceeds 20% (ecological stress) or the mean uncertainty weight exceeds 0.6 (high risk), the trigger delay is set to 15 minutes to match the Bayesian network iteration cycle in step S2. Only the uncertainty quantification layer (variational inference) and the data-driven layer (XGBoost) are updated, while the parameters of the mechanism constraint layer are retained to achieve rapid adaptation.
[0091] The offline training phase utilizes full data from the past 10 years, with a temporal resolution of daily hydrological-ecological data and hourly uncertain data, balancing completeness and granularity. It runs on an RTX A6000 GPU server, processing 5000 data points per batch. The dataset is stratified by life cycle stage and uncertainty level; for example, spawning period and high uncertainty samples account for 30%. A combined loss function with penalties for prediction bias and mechanistic bias is designed. Violations of ecological mechanism judgment conditions—water flow pulsation <0.5 times / hour or dissolved oxygen <8 mg / L during fish spawning period, and soil moisture content <15% during vegetation greening period—incur an additional 50% loss, strengthening mechanistic constraints. The AdamW optimizer minimizes the loss (learning rate 0.001, weight decay 0.01 to avoid overfitting), ensuring the model learns data patterns and ecological mechanisms simultaneously.
[0092] Specifically, step S6 is as follows:
[0093] Based on the baseline water demand (including confidence interval) output by the model trained in step S5, supply and demand are optimized in a coordinated manner. A formula for prioritizing ecological importance and uncertainty risk is established to achieve dynamic ranking. A multi-objective robust optimization function is established with the objectives of minimizing the water demand gap in sensitive areas (priority 1), minimizing water resource waste (priority 2), and maximizing the robustness of the scheme (priority 3). Three levels of constraints are set: water demand in sensitive areas at the river section level ≥ baseline value minus uncertainty lower limit; water supply at the sub-basin level ≥ total water demand in sensitive areas within the jurisdiction; total water supply at the basin level ≥ total water demand in all sub-basins, and the gap is ≤3% under the superposition of extreme drought (water supply reduced by 20%) and strong disturbance (water intake increased by 15%) scenarios.
[0094] An improved NSGA-III algorithm combined with reinforcement learning was used to solve the problem. The improvement involved optimizing the crossover operator and using an adaptive crossover probability of 0.4-0.8 to enhance solution diversity. NSGA-III was set to a population of 100 and 50 iterations to balance solution diversity and convergence, outputting a set of 3-5 Pareto optimal solutions. The algorithm was allocated at the watershed, sub-watershed, and river segment levels. At the watershed level, the ecological water replenishment effect data of sub-watersheds in the same period of the past 3 years were referenced, with evaluation indicators including fish spawning rate and vegetation survival rate. Quotas were prioritized for sub-watersheds with significant improvement in effect. At the sub-watershed level, the quotas were allocated to river segments. Sensitive areas were buffered by 5% more than the baseline water demand. Sensitive areas were defined as fish spawning grounds (historical spawning rate ≥50%) and core vegetation areas (chlorophyll content ≥50mg / m²). Ordinary river segments were buffered by 5% less than the baseline water demand. The buffer amount was referenced from the uncertainty interval in step S4, forming a multi-scale water replenishment scheme.
[0095] Multi-scale ecological water replenishment dynamic priority calculation formula:
[0096]
[0097] In the formula: This represents a water replenishment priority index, used to rank the water replenishment needs of different river sections / regions. The higher the value, the higher the priority for hydration;
[0098] The ecological importance weight ranges from 0 to 1 and is determined based on the ecological function of the river section, including spawning grounds and core vegetation areas. =0.9, ordinary river section =0.3;
[0099] This represents the ecological importance level value, ranging from 0 to 1, and... Supporting facilities, spawning grounds =1.0, Vegetation Zone =0.7, ordinary river section =0.3, quantifying the differences in ecological value among different regions;
[0100] This represents the uncertainty risk weight, with a value ranging from 0 to 1. It is positively correlated with the degree of environmental risk, and is associated with high-risk areas (such as areas with dense industrial water intake). =0.8, low-risk area =0.2;
[0101] This represents the uncertainty risk level value, ranging from 0 to 1, determined according to the uncertainty weight quantified in step S2. High uncertainty (weight...) =1.0, low uncertainty (weight) =0.4, quantifying the impact of environmental risk on water replenishment ranking.
[0102] Specifically, step S7 is as follows:
[0103] Based on the scaled water replenishment plan in step S6, a digital twin decision-making system is established. A three-dimensional visualization model is constructed using Unity3D and combined with GIS geographic data, with an accuracy of 1:10000 to ensure spatial positioning accuracy. The water replenishment plan is input into the twin model to simulate the ecological responses after water replenishment, such as fish spawning rate (simulation parameters are calibrated based on data from the HY-AS200 sensor in step S1) and vegetation survival rate (simulation step length is 1 hour, adapted to the uncertainty update cycle of step S4). If the simulated vegetation survival rate is lower than 60%, the system automatically returns to step S6 for re-optimization, forming a virtual-real closed loop, with a time consumption of ≤2 hours.
[0104] An automatic biological data feedback module is deployed to collect biological data in real time after water replenishment using the underwater camera (4K resolution, 25fps frame rate) and vegetation spectral sensor (400-1000nm band) already deployed in step S1. This data includes fish spawning count and chlorophyll content. The data is uploaded to the cloud via the MQTT protocol and stored in the cloud for 3 years. A dual-backup mechanism is used to ensure data security and traceability. The module is connected to the 5G transmission technology in step S1 to ensure real-time feedback. The ecological effect deviation rate is automatically calculated weekly using the formula: (actual ecological indicator value - simulated prediction value) / simulated prediction value × 100%. The ecological indicators include fish spawning rate and vegetation survival rate. For example, if the actual spawning rate is 15% lower than the prediction, the deviation value is input into the model training stage in step S5 to achieve ecological iteration without human intervention.
[0105] The interface links uncertainty and ecological impact. The left side displays the river section-level biological stress status, with abnormal spawning areas marked in red, based on biological monitoring data from step S1. The middle shows the distribution of uncertainty sources at the sub-basin level, with agricultural water diversion areas marked in yellow, corresponding to the uncertainty labels in step S1. The right side presents suggestions for adjusting the basin-level water replenishment plan, referring to the optimization results in step S6. The interface response time is ≤0.5 seconds, and it has the function of clicking to view the correlation chain of mechanism, uncertainty, and ecological effect.
[0106] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for predicting ecological water replenishment in rivers and lakes based on intelligent sensing, characterized in that: The specific steps of this method are as follows: Step S1: Set up data acquisition sensors in the watershed and label the sensors with ecological mechanism labels and uncertainty labels. Adjust the sampling frequency according to the uncertainty level, establish a multi-dimensional data linkage mechanism for biological anomaly triggering, and output the raw data. Step S2: Update the uncertainty weights using a Bayesian network, verify the validity of hydrological parameters using biological response data, remove outliers and fill in missing values, and standardize the units using Z-Score. Step S3: First calculate the importance of features, then select the core features, establish a three-level feature linkage matrix of river section, sub-basin, and basin, divide the feature time period according to the biological life cycle stage, align the hydrological and ecological data and label the associated tags; Step S4: Based on the core features selected in Step S3, a three-layer linkage model is established. The mechanism constraint layer incorporates the biological stress response sub-model. On the basis of the basic ecological water demand model during the fish spawning period, stress-corrected water demand is added to form a two-dimensional model of basic water demand and stress water demand. Results that do not conform to the mechanism are forcibly corrected. The data-driven layer adopts a combined model of LSTM and XGBoost. The LSTM is set to a 3-layer structure to capture the temporal data dependencies including water level and evapotranspiration. The XGBoost is set to a learning rate of 0.1 and a tree depth of 8 to fit the nonlinear relationship between soil moisture content and water demand. The core features of step S3 are input to output the baseline ecological water demand. The uncertainty quantification layer introduces variational inference, dynamically adjusting the water demand fluctuation range every hour based on biological anomaly data and uncertainty data, and finally outputs the baseline water demand, confidence interval, and ecological interpretation label. Step S5: Based on the model in Step S4, perform model training and transfer optimization, automatically perform incremental updates according to ecological and risk thresholds, and train the model offline so that the model can learn data patterns and ecological mechanisms simultaneously. Step S6: Prioritize water supply by combining ecological importance with uncertainty risk formula, establish multi-objective robust optimization function, use improved NSGA-III and reinforcement learning to solve 3-5 optimal solutions, and allocate water supply according to watershed, sub-watershed and river segment to form a multi-scale water replenishment scheme. Step S7: Establish a digital twin decision-making system to simulate the ecological response after water replenishment, set up an automatic biological data feedback module, and design an interface linking uncertainty and ecological impact.
2. The method for predicting river and lake ecological water replenishment based on intelligent sensing according to claim 1, characterized in that: The specific steps of S1 are as follows: Three monitoring sections were set up within the watershed. Each section was equipped with a sensing terminal and simultaneously labeled with ecological mechanism tags and uncertainty tags: underwater acoustic sensors labeled fish spawning period and upstream water fluctuations; soil moisture sensors labeled vegetation transpiration and agricultural emergency water diversion; water level and flow sensors labeled hydrological conditions and extreme weather; dissolved oxygen sensors labeled dissolved oxygen threshold for fish spawning and algal bloom interference; underwater cameras labeled fish spawning volume, egg and larval survival status and water turbidity interference; and vegetation spectral sensors labeled vegetation chlorophyll content, coverage and photosynthetic efficiency, realizing the binding of mechanism and risk association during the data collection stage. The sampling frequency is adjusted according to the level of uncertainty. When meteorological satellites detect extreme weather warnings or when government interfaces receive strong interference from industrial water intake, the sampling frequency of sensors in the corresponding areas is increased to 5 minutes / time, while the sampling frequency in non-interference areas is maintained at 1 hour / time. Data transmission uses EC-500 edge computing nodes paired with 5G slicing; When the underwater camera detects a sudden drop in fish spawning, a chlorophyll content below 35 mg / m² as monitored by the vegetation spectral sensor, a sudden drop in fish activity frequency, or abnormal vegetation chlorophyll content, it automatically triggers synchronous high-frequency data collection from related sensors, including upstream flow, dissolved oxygen, and surrounding water intake gates, forming a complete data chain of biological anomalies, hydrological parameters, and interference sources.
3. The method for predicting river and lake ecological water replenishment based on intelligent sensing according to claim 2, characterized in that: Step S2 is as follows: Based on the multi-source data from step S1, the data is processed according to the dual requirements of ecological mechanism and uncertainty. A Bayesian network is used to update the uncertainty weight in real time. The initial labeling is weak interference from agricultural water diversion. If the water diversion volume is detected to exceed the registered value by 50%, the weight is updated to 0.6, and the associated vegetation transpiration water demand data is marked as high uncertainty associated data. The validity of hydrological parameters is verified using biological response data. If the hydrological data shows a water flow pulsation frequency of 0.6 times / hour, but the underwater acoustic sensor deployed in step S1 detects that the fish spawning rate is less than 30%, it is determined that there is a mechanism deviation, triggering a second sampling and correction of the dissolved oxygen sensor. By constraining ecological thresholds and removing outliers using isolated forests, and by filling in missing values through mechanistic similarity interpolation, a dual-association dataset is finally formed through Z-Score standardization and unified dimensionality.
4. The method for predicting river and lake ecological water replenishment based on intelligent sensing according to claim 3, characterized in that: Step S3 is as follows: Based on the dual-association dataset from step S2, features are selected using a dual criterion of SHAP value and ecological mechanism. First, the importance of features is calculated through random forest, and then the influence of features on the uncertainty interval is analyzed through SHAP value. Irrelevant features are eliminated, and finally 15-20 core features are retained. A three-level feature linkage matrix was established for river sections, sub-basins, and watersheds. When the water demand for fish spawning at the river section level is correlated with the water storage capacity of reservoirs at the sub-basin level, an uncertainty transmission coefficient was introduced. When the sub-basin is in drought with high uncertainty, the coefficient was set to 1.2, which means the water demand was increased by 20%. At the same time, feature time periods were divided according to the biological life cycle stage. The Pandas library of Python was used to slice and align the hydrological and ecological data according to the key time periods and label the correlation.
5. The method for predicting river and lake ecological water replenishment based on intelligent sensing according to claim 4, characterized in that: Step S5 is as follows: Based on the three-layer linkage model in step S4, the core mechanism parameters are fixed during cross-scale migration, and the stress coefficient of fish and vegetation transpiration coefficient at the river section level are used as anchor parameters. When migrating to the sub-basin, only the uncertainty transmission coefficient of the data-driven layer is finely adjusted to ensure the consistency of biological mechanisms in the same basin. When the monitoring shows that the wilting rate of aquatic plants exceeds 20% or the mean of uncertainty weight exceeds 0.6, the trigger delay is set to 15 minutes, and only the uncertainty quantification layer and the data-driven layer are updated, while the parameters of the mechanism constraint layer are retained. The offline training phase uses full data from the past 10 years, including hydrological, ecological, and uncertain data, and runs on an RTX A6000 GPU server. The dataset is divided into layers according to life cycle stage and uncertainty level. A comprehensive loss function with penalty terms for prediction bias and mechanism bias is designed, and the loss is minimized by the AdamW optimizer to ensure that the model learns data patterns and ecological mechanisms simultaneously.
6. The method for predicting river and lake ecological water replenishment based on intelligent sensing according to claim 5, characterized in that: Step S6 is as follows: Based on the baseline water demand output by the model trained in step S5, supply and demand are optimized in a coordinated manner. A formula for prioritizing ecological importance and uncertainty risks is established to achieve dynamic ranking. A multi-objective robust optimization function is established with the objectives of minimizing the water demand gap in sensitive areas, minimizing water waste, and maximizing the robustness of the scheme. Three levels of constraints are set: water demand in sensitive areas at the river section level ≥ baseline value minus uncertainty lower limit; water supply at the sub-basin level ≥ total water demand in sensitive areas within the jurisdiction; total water supply at the basin level ≥ total water demand in all sub-basins, and the gap ≤ 3% under the scenario of extreme drought and strong disturbance. An improved NSGA-III algorithm combined with reinforcement learning is used to solve the problem, outputting a set of 3-5 Pareto optimal solutions. The schemes are allocated at the basin, sub-basin, and river section levels. At the basin level, the water supply quota for the sub-basin is determined by referring to the ecological water replenishment effect data of the same period in the past 3 years. At the sub-basin level, the quota is allocated to the river section to form a multi-scale water replenishment scheme.
7. The method for predicting river and lake ecological water replenishment based on intelligent sensing according to claim 6, characterized in that: Step S7 is as follows: Based on the scaled water replenishment plan in step S6, a digital twin decision-making system is established. A three-dimensional visualization model is constructed using Unity3D and combined with GIS geographic data. The water replenishment plan is input into the twin model to simulate the ecological response, including fish spawning rate and vegetation survival rate after water replenishment. If the simulated vegetation survival rate is less than 60%, the system automatically returns to step S6 for re-optimization, forming a virtual-real closed loop. Deploy an automatic biological data feedback module to collect biological data in real time after water replenishment using the underwater camera and vegetation spectral sensor deployed in step S1. The data is uploaded to the cloud via the MQTT protocol. The ecological effect deviation rate is automatically calculated weekly, and the deviation value is input into the model training stage in step S5. The interface links uncertainty and ecological impact. The left side displays the biological stress status at the river section level, the middle shows the distribution of uncertainty sources at the sub-basin level, and the right side presents suggestions for adjusting the watershed-level water replenishment plan. It also has the function of clicking to view the correlation chain between mechanism, uncertainty, and ecological effect.
Citation Information
Patent Citations
Reservoir ecological flow configuration method based on multi-section water transfer and multi-target fish collaborative protection
CN119809263A
Coastal water level monitoring method, medium and system based on dynamic threshold algorithm
CN120593866A
