A Method and System for Estimating Irrigation District Water Consumption Based on Adaptive Dynamic Feature Fusion

By using dynamic feature fusion and incremental training of a deep LSTM model, the problems of dynamic changes and periodic drift in irrigation water data are solved, achieving high-precision short-term prediction and long-term trend capture, thus improving the scientific nature of irrigation scheduling and water use efficiency.

CN122114490BActive Publication Date: 2026-11-17SHANDONG HAOZHI INFORMATION TECHNOLOGY CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202610206912.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-02-12
Publication Date
2026-11-17
Estimated Expiration
2046-02-12

AI Technical Summary

Technical Problem

Existing technologies cannot effectively capture the dynamic changes, periodic drift behavior, and non-fixed time-delay response patterns in irrigation water data, resulting in prediction lag, trend misjudgment, and performance degradation after long-term deployment. They cannot simultaneously meet the needs of short-term scheduling and long-term trends, and are difficult to adapt to data distribution drift caused by changes in irrigation facilities or climate anomalies.

Method used

By constructing an adaptive dynamic feature fusion method based on a deep LSTM model, utilizing dynamic lag embedding matrix and periodic feature fusion, combined with multi-scale feature pyramid output and incremental training, the model's adaptability to sudden changes in meteorological response and its long-term adaptability are improved, achieving high-precision short-term forecasting and robust long-term trend capture.

Benefits of technology

It significantly improves the scientific nature of irrigation scheduling and water use efficiency, can adaptively handle complex spatiotemporal changes, and enhances the model's ability to respond to emergencies and adapt to adjustments in irrigation district structure.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122114490B_ABST
    Figure CN122114490B_ABST
Patent Text Reader

Abstract

This invention relates to the fields of artificial intelligence and data processing technology, specifically to a method and system for estimating irrigation water consumption based on adaptive dynamic feature fusion. The method involves: real-time collection of irrigation water consumption data; preprocessing the collected data; constructing a dynamic lag embedding matrix based on historical water consumption information, and then fusing it with periodic features to obtain historical fused features; constructing an irrigation water consumption estimation model based on a deep LSTM model, inputting the preprocessed data and historical fused features into the model for training until preset conditions are met, resulting in a trained model; incrementally training the trained model; and inputting newly collected irrigation water consumption data and meteorological data into the incrementally trained model to estimate irrigation water consumption. This invention can solve the problems of prediction bias and inaccuracy caused by dynamic data, thereby improving the scientific nature of irrigation scheduling and the robustness of water use efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data processing technology, and in particular to a method and system for estimating irrigation water consumption based on adaptive dynamic feature fusion. Background Technology

[0002] Agricultural irrigation water estimation, as an important component of the smart water conservancy system, provides crucial data support for the system. The deep integration of agricultural irrigation water estimation and smart platforms is evolving from a simple technological overlay to a systemic reconstruction. Through accurate water consumption forecasting, the smart water conservancy system can better allocate and schedule water resources, optimize irrigation plans, and reduce water waste.

[0003] With the gradual deployment of smart water meters, remote sensing equipment, and IoT technology in agricultural irrigation areas, massive amounts of water consumption monitoring data have been acquired and centrally processed in real time. However, these data exhibit significant non-stationarity, strong periodicity, and response lag characteristics. They are influenced by multiple factors such as rainfall, evaporation, planting structure, and agricultural activities, and are also subject to nonlinear disturbances caused by sudden events (such as rainstorms and droughts). Therefore, relying solely on traditional time series models (such as ARIMA, SVR, and Prophet) or standard Long Short-Term Memory (LSTM) networks cannot effectively capture the dynamic changes, periodic drift behavior, and non-fixed time-delay response patterns in irrigation water data. This can easily lead to prediction lag, trend misjudgment, or seasonal bias, seriously affecting the scientific nature of irrigation scheduling and the optimization of water use efficiency.

[0004] Existing technologies suffer from the following objective problems: Traditional sliding window methods use a fixed step size, ignoring the dynamic changes in water use response with weather and seasons, and cannot accurately model the actual delay relationship between rainfall and irrigation; existing methods often do not explicitly model intraday and monthly cyclical structures, making it difficult for models to understand the impact of crop planting structure and the periodicity of agricultural activities on water use behavior; standard LSTM structures suffer from forget gate response delays in handling sudden events, and most models can only output single-granularity prediction results, failing to simultaneously address short-term scheduling and long-term trend requirements; conventional models cannot detect and adapt to data distribution drift caused by changes in irrigation facilities or climate anomalies, lack incremental training and parameter alignment mechanisms, resulting in severe performance degradation after long-term deployment.

[0005] Therefore, this invention proposes a method and system for estimating irrigation water consumption based on adaptive dynamic feature fusion to solve the above problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention develops a method and system for estimating irrigation water consumption based on adaptive dynamic feature fusion. This invention improves the accuracy of model prediction results by training a deep LSTM model used for water consumption estimation and by incrementally training the deep LSTM model.

[0007] On the one hand, the technical solution of this invention to solve the technical problem is a method for estimating irrigation water consumption based on adaptive dynamic feature fusion, which includes the following steps: S1. By deploying smart water meters, flow meters and remote sensing monitoring equipment in the irrigation area at all levels, the water consumption data of the irrigation area is collected in real time, and then the collected data is transmitted to the central database for storage through the Internet of Things terminal. S2. Read the collected water consumption data from the central database and perform data cleaning, time alignment and normalization. S3. Construct a dynamic lag embedding matrix based on historical water consumption information, and then fuse it with periodic features to obtain historical fused features; S4. Construct an agricultural irrigation water consumption estimation model based on a deep LSTM model. Input the preprocessed data and historical fusion features into the model for training until the model meets the preset conditions and then stop training to obtain a trained agricultural irrigation water consumption estimation model. S5. Incremental training is performed on the trained irrigation water consumption estimation model for the irrigation area. S6. Input the newly collected quantitative and meteorological data of the irrigation area into the incrementally trained model to estimate the agricultural irrigation water consumption of the irrigation area.

[0008] S3 is as follows: S3.1 Constructing a dynamic lag embedding matrix for historical water consumption data: Based on the time-varying mutual information between irrigation district water consumption data and meteorological factors, a dynamic lag embedding matrix is ​​constructed. First, the time-varying mutual information between historical water consumption data and meteorological factors is calculated to determine the optimal lag step size for each time point. Then, according to the determined optimal lag step size, the water consumption observation values ​​of each irrigation district at different lag time points are extracted from the historical data. The water consumption observation values ​​are arranged in the order of irrigation district and lag step size to form a dynamic lag embedding matrix. Each row of the matrix corresponds to the water consumption of an irrigation district at different lag time points, and each column corresponds to the water consumption of all irrigation districts at a specific lag time point. The dynamic lag embedding matrix of historical water consumption data is represented as follows: , in, For dynamic hysteresis embedding matrix; For irrigation district 1 in time Water consumption observations; For irrigation district 1 in time Water consumption observations; For irrigation area Na in time Water consumption observations; For irrigation area Na in time Water consumption observations; Indexed by the current time; This represents the maximum lag step size. The minimum lag step size; The maximum lag order; Number of irrigation districts; S3.2, Perform periodic feature fusion of irrigation district water consumption data: In the process of constructing periodic features, time embedding features in two dimensions—daily and monthly—are extracted based on the current time index. The daily period embedding represents the phased changes in daily irrigation activities, while the monthly period embedding represents the macro-adjustment trend of planting structure. Then, the two periodic features are input into a weight transformation function to calculate the periodic gating vector. The periodic gating vector is used to control the fusion ratio between historical water consumption features and periodic features, thereby generating a periodically guided feature vector. The fused feature vector is calculated using the following formula: , , , , in, This represents the daily periodic eigenvector; This represents the monthly cycle feature vector; This involves concatenating vectors. It is a periodic gated vector; Use the Sigmoid activation function; This is element-wise multiplication; The characteristics of historical integration after fusion; This is the weight matrix for the daily cycle characteristics; This is the weight matrix for the monthly cycle characteristics; The weight matrix for the periodic feature transformation; This indicates vector concatenation; For the dynamic hysteresis embedding matrix in time The input vector.

[0009] S4 is as follows: A deep LSTM model is constructed to estimate agricultural irrigation water consumption in irrigation districts. The model is a deep network structure containing multiple LSTM units. Each LSTM unit extracts temporal features at different time scales. The first LSTM unit receives the fused feature vector as input and gradually abstracts higher-order temporal dependencies through stacked hidden layers. The last LSTM unit is connected to the multi-scale feature pyramid output layer to fuse the daily and weekly hidden states into the final predicted value. The preprocessed data and the fused historical features are input into the model for training. First, the trainable parameters of the model are initialized, the weight matrix and bias term of the LSTM unit are initialized with a normal distribution, and the initial variance and forget gate bias term are set. Then, the model is trained by improving the forward propagation process, constructing a multi-scale feature pyramid output, calculating the loss function, performing gradient-oriented pruning optimization, and performing memory cell sparsity regularization. Mini-batch gradient descent is used to train the model for each iteration. Each batch samples time-series segments from the preprocessed data and inputs them into the deep LSTM model. After forward propagation generates predicted values, the parameter update amount is calculated using the calculated loss function and the gradient of the loss function with respect to the trainable parameters. During backpropagation, the model parameters are updated in conjunction with the gradient pruning strategy. The learning rate of the deep LSTM model adopts an exponential decay strategy until the loss function does not decrease significantly in multiple consecutive iterations or the difference between the training error and the validation error is lower than a preset threshold. Once the model meets the preset conditions, training stops, resulting in a well-trained irrigation water estimation model for the irrigation district.

[0010] The specific steps for improving the forward propagation process of the deep LSTM model are as follows: The water consumption input at the current time step is jointly projected onto the hidden state at the previous time step, and a normalized gradient factor is constructed by combining the difference between the water consumption at the current time step and the previous time step. The normalized gradient factor is injected into the forget gate in a weighted form to obtain the improved forget gate. At the same time, during the state update process, the previous state and the candidate state are used to determine the current memory cell state through the gate controller.

[0011] The specific steps for constructing the multi-scale feature pyramid output are as follows: First, the input at the current time is encoded using a standard LSTM unit to obtain the daily granular hidden state. Then, max pooling is performed on the daily granular state every 7 time steps to extract the most salient features within a week, forming the weekly granular hidden state. Finally, in the output stage, the current daily granular state and its corresponding weekly granular state are weighted and fused to output the final water consumption prediction result.

[0012] The calculation process of the loss function is as follows: Based on the deviation between the prediction error and the actual error, an error term is constructed for each time step. A predefined error threshold is used to determine whether the error needs to be penalized. If the error is greater than the error threshold, the seasonal importance factor is used to enhance its weight in the loss function. Otherwise, the basic error weight is maintained. Finally, the loss terms of all time steps are weighted and averaged, and combined with the regularization term to obtain the overall loss.

[0013] The specific process of gradient-oriented clipping optimization is as follows: During the model parameter update phase, the gradient vector at the current time step is first recorded, and the average direction is calculated based on the gradient values ​​of several historical steps. If the current gradient value exceeds the clipping threshold and the direction is highly consistent with the historical gradient direction, the current gradient size is scaled according to the historical average direction to perform directed clipping of excessively large gradients.

[0014] The specific process of sparse regularization of memory units is as follows: The LSTM memory states are divided into several groups. The L2 norm of each group of memory cells is calculated, and a weight term based on the group size is applied to form a structured regularization penalty term. The calculation formula is as follows: , in, For regularization terms; Number of groups for memory units; Grouped index; Represents the norm symbol; For the first Group memory units; Indicates the first The number of group memory units; The sparsity strength is denoted by .

[0015] S5 is detailed below: S5.1 Define the incremental training trigger conditions: Incremental training is automatically triggered when the distribution difference between the newly added water consumption data and the historical data exceeds the threshold, or when the prediction error of the irrigation district agricultural irrigation water consumption estimation model on the latest data continues to be higher than the acceptable threshold. S5.2 Incremental Online Timing Alignment: During incremental training, an incremental online temporal alignment method is adopted. When using new data for incremental model training, the Wasserstein-1 bulldozer distance between the distributions of the new and old sample data is first calculated. The Wasserstein-1 distance is then combined with the loss function on the new data for incremental training, and the current iteration parameters are guided to move closer to the historical parameters. The degree of closeness is adjusted according to the distribution differences. The calculation formula is as follows: , , in, for Wasserstein-1 distance between the distributions of new and old data at time points; The Wasserstein-1 distance calculation function; This represents the input data; Distribution of old data; For the new data distribution; This is the set of trainable parameters for the updated LSTM model; This is the set of trainable parameters for the old LSTM model; This is the set of trainable parameters for the LSTM model in the current iteration during incremental training. Incremental training learning rate; For gradient operators; The loss function for incremental training on new data; Alignment strength coefficient; S5.3, Evaluate the incremental training effect: After incremental training is completed, the performance of the irrigation district agricultural irrigation water consumption estimation model is evaluated using an independent test set. If the prediction accuracy recovers to the preset standard and the performance is balanced on the old and new data, then update the parameters of the irrigation district agricultural irrigation water consumption estimation model; otherwise, adjust the alignment strength coefficient and retrain.

[0016] On the other hand, the present invention also provides an irrigation water consumption estimation system based on adaptive dynamic feature fusion, including a module for executing processing instructions for each step in the irrigation water consumption estimation method based on adaptive dynamic feature fusion. Sensing and data acquisition module: Collects water consumption data of the irrigation area through equipment deployed in various levels of canals in the irrigation area, and transmits the collected data to a central database for storage via the LoRaWAN protocol; Data preprocessing module: Reads water consumption data collected from the central database, performs data cleaning, time alignment and normalization, and constructs a high-quality training dataset; Feature fusion engine module: includes feature enhancement module, dynamic lag embedding module and periodic gating fusion module; real-time meteorological covariates are injected into historical water consumption information to construct dynamic lag embedding matrix, and then fused with periodic features to obtain historical fused features; Irrigation water consumption estimation module for irrigation districts: The preprocessed data and historical fusion features are input into the irrigation district agricultural irrigation water consumption estimation module for training until the model meets the preset conditions and then training stops. Incremental learning module: Incrementally trains the agricultural irrigation water consumption estimation module in the irrigation area after training; Application Decision Module: Input the new real-time collected irrigation district quantity data and meteorological data into the irrigation district agricultural irrigation water consumption estimation module after incremental training to estimate the irrigation district agricultural irrigation water consumption.

[0017] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention effectively models non-fixed time-lag features by calculating the time-varying mutual information between water consumption and meteorological factors, adaptively determining the optimal lag step size, and dynamically constructing historical input sequences. This improves the model's adaptability to sudden meteorological responses and solves the problem of distortion of short-term abrupt features due to the strong periodicity of irrigation water consumption data, which cannot adapt to dynamic data changes. By combining daily and monthly periodic embeddings and using Sigmoid gating to dynamically adjust the fusion ratio of periodic information and historical states, this invention enhances the model's ability to characterize changes in planting structure and irrigation operation patterns, avoiding the problem of mixing periodic and time series data and making it difficult to intuitively display the components corresponding to different periods. Furthermore, by injecting a normalized gradient factor into the forget gate to construct a mutation-sensitive forget gate, combined with multi-scale feature pyramid output, this invention avoids the problem of peak prediction lag, thus significantly improving the model's fine-grained response and trend capture capabilities in sudden water use events. It can simultaneously perform high-precision short-term prediction and robust long-term trend capture. Finally, based on Wasserstein distance to determine the difference in distribution between old and new data, this invention combines distribution alignment and loss joint optimization to achieve online incremental training, improving the model's long-term adaptability to irrigation structure adjustments and climate change.

[0018] Therefore, this invention models time features by using dynamic lag of time-varying interactive information and periodic gating, designs a dynamic architecture by using gradient factor forget gate and constructing multi-scale pyramid output, trains the model's stability by using directional gradient pruning and grouped sparsity regularization, and improves the model's long-term adaptability by using Wasserstein distance increment. This invention constructs an irrigation water consumption estimation model for irrigation districts that can adaptively handle complex spatiotemporal changes, thereby improving the scientific nature of irrigation scheduling and the robustness of water use efficiency. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0021] Figure 2 A comparison chart showing the prediction effects of different lag compensation methods.

[0022] Figure 3This is a comparison chart showing the accuracy of the method of the present invention in estimating irrigation water consumption with existing methods based on the root mean square error.

[0023] Figure 4 This is a comparison chart showing the accuracy of the method of the present invention in estimating irrigation water consumption with existing methods by means of mean absolute error.

[0024] Figure 5 This is a comparison chart showing the accuracy of estimating irrigation water usage using the method of the present invention and existing methods, based on the coefficient of determination.

[0025] Figure 6 This is a comparison chart showing the estimated irrigation water consumption when the method of this invention and existing methods are used to handle sudden events.

[0026] Figure 7 This is a comparison chart of the training loss of the method in this invention and existing methods in multiple training rounds.

[0027] Figure 8 This is a comparison chart of the prediction error distribution of the method in this invention and existing methods in different seasons.

[0028] Figure 9 This is a performance comparison chart of the incremental training method and the fixed window fine-tuning method in this invention. Detailed Implementation

[0029] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0030] Example 1 like Figure 1 As shown, a method for estimating irrigation district water consumption based on adaptive dynamic feature fusion includes the following steps: S1. By deploying smart water meters, flow meters and remote sensing monitoring equipment in the irrigation area at all levels, the water consumption data of the irrigation area is collected in real time, and then the collected data is transmitted to the central database for storage through the Internet of Things terminal. Data acquisition covers historical periods and real-time update streams, ensuring the continuity of time series; the acquisition frequency is set according to irrigation management needs, usually using hourly or daily granularity, which can fully capture the dynamic changes in water consumption. S2. Read the collected water consumption data from the central database and perform data cleaning, time alignment and normalization. Data cleaning includes identifying and removing missing values, outliers, and noise interference caused by equipment failure or transmission anomalies; to address the potential timestamp offset issues in data from different irrigation districts, linear interpolation or nearest neighbor filling is used for time alignment to ensure that all irrigation district data have a unified time reference; by normalizing the cleaned data, dimensional differences are eliminated, and water consumption and meteorological factors are scaled to the [0,1] interval. S3. Construct a dynamic lag embedding matrix based on historical water consumption information, and then fuse it with periodic features to obtain historical fused features; S4. Construct an agricultural irrigation water consumption estimation model based on a deep LSTM model. Input the preprocessed data and historical fusion features into the model for training until the model meets the preset conditions and then stop training to obtain a trained agricultural irrigation water consumption estimation model. S5. Incremental training is performed on the trained irrigation water consumption estimation model for the irrigation area. S6. Input the newly collected quantitative and meteorological data of the irrigation area into the incrementally trained model to estimate the agricultural irrigation water consumption of the irrigation area.

[0031] In a specific implementation, S3 is as follows: S3.1 Constructing a dynamic lag embedding matrix for historical water consumption data: Because irrigation water consumption data is highly periodic, cycling daily or monthly, and has a delayed response characteristic, the conventional sliding window method, with its fixed lag step, ignores the non-fixed time lag between meteorological events and water consumption response. This makes it unable to adapt to the dynamic changes in irrigation response time after rainfall with the seasons, resulting in distorted capture of short-term abrupt change characteristics. Therefore, this invention constructs a dynamic lag embedding matrix based on the time-varying mutual information between irrigation district water consumption data and meteorological factors. First, the time-varying mutual information between historical water consumption data and meteorological factors is calculated to determine the optimal lag step size for each time point. The lag step size reflects the dynamic time-lag relationship between meteorological events and water consumption responses. Then, based on the determined optimal lag step size, water consumption observations for each irrigation district at different lag time points are extracted from historical data. The water consumption observations are arranged in the order of irrigation district and lag step size to form a dynamic lag embedding matrix. Each row of the matrix corresponds to the water consumption of an irrigation district at different lag time points, and each column corresponds to the water consumption of all irrigation districts at a specific lag time point. The dynamic lag embedding matrix of historical water consumption data is represented as follows: , in, For dynamic hysteresis embedding matrix; For irrigation district 1 in time Water consumption observations; For irrigation district 1 in time Water consumption observations; For irrigation area Na in time Water consumption observations; For irrigation area Na in time Water consumption observations; Indexed by the current time; Set the maximum lag step size. ; Set the minimum lag step size. ; To set the maximum lag order, ; Number of irrigation districts; It should be noted that the maximum and minimum lag step sizes are determined based on mutual information calculations. Specifically, they are first based on a predefined maximum allowable lag range, such as... Every day, all feasible lag steps are enumerated. Then, for each lag step, the mutual information between the water consumption at that lag time point and the meteorological factor at the current time is calculated. Finally, the value with the largest mutual information value is selected as the optimal lag step for the current input, which is used to dynamically construct the historical embedding sequence.

[0032] S3.2, Perform periodic feature fusion of irrigation district water consumption data: Water consumption exhibits a bimodal characteristic during daily irrigation periods, and monthly changes in planting structure cause cyclical patterns to drift. Conventional LSTMs mix cyclical features with time series inputs, making it difficult to explicitly model the daily or monthly cyclical components. Therefore, in the process of constructing periodic features, time embedding features in two dimensions—daily and monthly—are extracted based on the current time index. The daily period embedding represents the phased changes in daily irrigation activities, while the monthly period embedding represents the macro-adjustment trend of the planting structure. Then, the two periodic features are input into a weight transformation function to calculate a periodic gating vector. This periodic gating vector controls the fusion ratio between historical water consumption features and periodic features, thereby generating a periodically guided feature vector. The fused feature vector is calculated using the following formula: , , , , in, It is a daily cycle feature vector that encodes the daily irrigation cycle pattern and represents the phased changes in daily irrigation activities; The monthly periodic feature vector encodes the pattern drift caused by changes in the monthly planting structure, and characterizes the pattern drift caused by changes in the monthly planting structure. This involves concatenating vectors. It is a periodic gate vector, generated by the Sigmoid activation function, which controls the fusion ratio between historical water consumption features and periodic features; Use the Sigmoid activation function; This is element-wise multiplication; The merged historical characteristics include historical water consumption information and periodic information; The weight matrix represents the daily periodic characteristics and consists of trainable parameters. It is trained and updated using backpropagation based on gradient descent. The weight matrix represents the monthly cycle features and consists of trainable parameters. It is trained and updated using backpropagation based on gradient descent. The weight matrix is ​​the periodic feature transformation, which is a trainable parameter. It is trained and updated using the error backpropagation method based on gradient descent. This indicates vector concatenation; For the dynamic hysteresis embedding matrix in time The input vector.

[0033] It should be noted that the periodic gate vector It is a periodic gated vector generated through Sigmoid activation, and the value of each dimension is in The interval between these intervals indicates the fusion strength of the periodic features in that dimension. Represents the dynamic lag embedding matrix in time The residual control weights of the input vector for that dimension, when When an element is close to 1, it indicates that more historical water consumption characteristics are preserved; when When an element is close to 0, it indicates the introduction of more periodic characteristics; therefore... The weights of the periodic features are adjusted to ensure that the fused feature vector contains both historical water consumption information and periodic information.

[0034] In a specific implementation, S4 is as follows: A deep LSTM model is constructed to estimate agricultural irrigation water consumption in irrigation districts. The model is a deep network structure containing multiple LSTM units. Each LSTM unit extracts temporal features at different time scales. The first LSTM unit receives the fused feature vector as input and gradually abstracts higher-order temporal dependencies through stacked hidden layers. The last LSTM unit is connected to the multi-scale feature pyramid output layer to fuse the daily and weekly hidden states into the final predicted value. The preprocessed data and the fused historical features are input into the model for training. First, the trainable parameters of the model are initialized. The weight matrix and bias term of the LSTM unit are initialized with a normal distribution. The initial variance and forget gate bias term are set. The initial variance is set to a small value to avoid gradient explosion. The initial variance is set to 0.0001. The forget gate bias term is set to a large positive value to ensure that the model tends to retain historical states in the early stage of training and avoids discarding effective information too early. The forget gate bias term is set to 0.5. Then, the model is trained by improving the forward propagation process, constructing a multi-scale feature pyramid output, calculating the loss function, performing gradient-oriented pruning optimization, and performing memory unit sparsity regularization. Mini-batch gradient descent is used to train the model for each iteration. Each batch samples time-series segments from the preprocessed data and inputs them into the deep LSTM model. After forward propagation generates predicted values, the parameter update amount is calculated using the calculated loss function and the gradient of the loss function with respect to the trainable parameters. During backpropagation, the model parameters are updated in conjunction with the gradient pruning strategy. The learning rate of the deep LSTM model adopts an exponential decay strategy, with rapid convergence in the early stage and fine-tuning in the later stage. The initial learning rate is set to 0.001 until the loss function does not decrease significantly in consecutive iterations or the difference between the training error and the validation error is lower than a preset threshold. Once the model meets the preset conditions, training stops, and a well-trained irrigation water estimation model for the irrigation district is obtained.

[0035] The conditions for stopping model iteration are set as follows: the change in the loss function is less than 0.001 for five consecutive training cycles; or the difference between the training error and the validation error is lower than a preset threshold of 0.05, which can prevent overfitting.

[0036] In a specific implementation, the operation of the improved forward propagation process of the deep LSTM model is as follows: Water consumption data in irrigation districts often exhibit both abrupt and gradual changes, such as the coexistence of drought-resistant days and rainy seasons. Conventional LSTM forget gates are slow to respond to abrupt events, which can easily lead to lag in peak prediction. Therefore, this invention improves the forward propagation process of the deep LSTM model by jointly projecting the water consumption input at the current time step with the hidden state at the previous time step, and constructing a normalized gradient factor by combining the difference between the water consumption at the current time step and the previous time step. The normalized gradient factor is injected into the forget gate in a weighted form to obtain an improved forget gate, thereby enhancing the LSTM's ability to respond to sudden water consumption patterns. At the same time, during the state update process, the previous state and the candidate state are used to determine the current memory cell state through the gate controller. The calculation formula for the above operation is as follows: , , , in, To improve the forget gate, a gradient-enhanced forget gate is characterized by normalization. The real-time water consumption gradient injects abrupt change information, enabling the forget gate to respond quickly to irrigation events; Here is the weight matrix for the forget gate; for The hidden state at time step 1, for the first layer of the deep LSTM model, has the fused feature vector as its input. ; For the bias term of the forget gate; The normalized gradient factor represents the normalized value of the difference between the water consumption at the current time and the water consumption at the previous time. for Water consumption at any given time; for Water consumption at any given time; This is the mutation sensitivity coefficient, used to adjust the weight of mutation information. ; This is a function to find the maximum value. For small constants, such as 0.0001; The current memory cell state represents long-term dependent information; This represents the state of the memory unit at the previous moment; It is the hyperbolic tangent function; The input gate vector; This is the weight matrix for the candidate memory states.

[0037] In a specific implementation, the operation of constructing the multi-scale feature pyramid output is as follows: Irrigation decisions rely on daily water consumption and weekly trends, which have multi-scale information characteristics. Conventional LSTM outputs cannot simultaneously satisfy high-precision short-term prediction and robust long-term trend capture. Therefore, this invention constructs a multi-scale feature pyramid output layer. First, the input at the current time is encoded using a standard LSTM unit to obtain the daily granular hidden state. Then, max pooling is performed on the daily granular state at every 7 time steps to extract the most significant features within a week and form the weekly granular hidden state. Finally, in the output stage, the current daily granular state and its corresponding weekly granular state are weighted and fused to output the final water consumption prediction result. The calculation formula for the above operation is as follows: , , , in, This represents the daily granularity hidden state, characterizing short-term temporal dynamics; For the forward propagation function of the daily granularity LSTM cell; This is a weekly granular feature, representing long-term trend information; It is the max pooling function; For the first Heaven to the First The daily hidden state sequence of the day; This is a predicted water consumption value; The weight matrix for daily granular output is a trainable parameter, trained and updated using backpropagation based on gradient descent. The weight matrix is ​​the weekly granular output, which consists of trainable parameters. It is trained and updated using backpropagation based on gradient descent. In order to keep pace with time The weekly granularity features of the corresponding weekly index.

[0038] It should be noted that the LSTM model uses forward propagation to process the fused feature vectors. Temporal modeling is performed, and a fine-grained state for the current moment is generated through an improved forget gate and state update mechanism, including a daily-granular hidden state for storing short-term temporal dynamics. and the current memory cell state used to carry long-term dependency information Multi-scale feature pyramids are based on daily granularity hidden states. The final prediction result is generated through pooling and fusion operations.

[0039] In a specific implementation, the calculation process of the loss function is as follows: In real-world scenarios, water consumption prediction errors have a more significant impact during peak irrigation seasons than during off-seasons. Conventional machine learning models typically treat all errors equally using mean squared error, ignoring seasonal differences. Therefore, when calculating the loss function, this invention constructs an error term for each time step based on the deviation between the prediction error and the actual error, and determines whether the error needs to be penalized by combining a predefined error threshold. If the error is greater than the error threshold, the seasonal importance factor is combined to enhance its weight in the loss function; otherwise, the basic error weight is maintained. Finally, the loss terms of all time steps are weighted and averaged, and the overall loss is obtained by combining the regularization term. The formula for calculating the loss function is as follows: , , in, The loss function characterizes the overall measure of the model's prediction error; This represents the total number of time steps. For regularization terms; This is a dynamic weighting factor used to adjust the error weights at different time steps; This is a predicted water consumption value; This represents the actual water consumption. As a seasonal importance factor, it characterizes the sensitivity of different seasons to errors. During peak season, it sets... During the off-season, set ; This is the error threshold, used to determine whether an error needs to be penalized. It is 0.2 times the maximum value among all actual water consumption values; The penalty intensity is used to adjust the weight of strong penalties. .

[0040] In a specific implementation, the gradient-oriented pruning optimization process is as follows: Water consumption data contains a large number of stable segments with small gradients and a small number of abrupt segments with large gradients, which can easily lead to optimization oscillations during LSTM training. In conventional techniques, global gradient clipping is difficult to adapt to non-stationary gradient distributions, and abrupt segments are prone to divergence. Therefore, this invention performs gradient-oriented pruning optimization. During the model parameter update stage, the gradient vector of the current time step is first recorded, and the average direction is calculated based on the gradient values ​​of several historical steps. If the current gradient value exceeds the pruning threshold and the direction is highly consistent with the historical gradient direction, the current gradient size is scaled according to the historical average direction to prune the excessively large gradient, thereby balancing the training stability of the model between the abrupt change segment and the stable segment.

[0041] The calculation formula for the above operation is as follows: , , , in, The current gradient represents the gradient of the loss function with respect to the parameters; This represents the gradient of the loss function with respect to the parameters. This is the set of trainable parameters for the LSTM model. The gradient after clipping; for Norm; The average gradient direction is preset for the historical time steps; Set the historical gradient window size. ; This is the clipping threshold, used to determine if the gradient is too large. ; The cosine similarity between vectors is calculated to determine the consistency of gradient directions; This is a directional consistency threshold used to determine whether the gradient directions are highly consistent. ; Let be the historical gradient vector, representing the gradient at the th . The gradient value at the time step.

[0042] In a specific implementation, the process of sparse regularization of memory units is as follows: Irrigation water usage is affected by many factors, but only a few dominant factors play a role in a specific period. Conventional techniques use standard L2 regularization to regularize the memory cells of LSTM, but cannot induce structured sparsity of memory cells. Therefore, this invention divides the LSTM memory states into several groups. By calculating the L2 norm of each group of memory units and applying a weight term based on the group size, a structured regularization penalty term is formed. This effectively identifies memory unit groups that are irrelevant to the current prediction task, guiding them to close as a whole, thereby improving the model's generalization ability and structural simplicity. The calculation formula is as follows: , in, This is a regularization term used to apply a sparsity penalty; The number of memory units is used to group the memory units, representing the grouping situation of the memory units; This is a grouped index used to identify different groups of memory units; Represents the norm symbol; For the first Group memory units represent the memory state of a specific group; Indicates the first The number of group memory units; Sparsity strength is used to adjust the weights of the sparsity penalty. .

[0043] In a specific implementation, S5 is as follows: S5.1 Define the incremental training trigger conditions: Incremental training is automatically triggered when the distribution difference between the newly added water consumption data and the historical data exceeds the threshold, or when the prediction error of the irrigation district agricultural irrigation water consumption estimation model on the latest data continues to be higher than the acceptable threshold. S5.2 Incremental Online Timing Alignment: During the data collection process of irrigation district water consumption, the historical data distribution may drift due to the renovation of irrigation district infrastructure. When training the model incrementally, the fixed window fine-tuning in conventional methods is difficult to adapt to the gradual changes in data distribution. During incremental training, an incremental online temporal alignment method is adopted. When using new data for incremental model training, the Wasserstein-1 bulldozer distance between the distributions of the new and old sample data is first calculated. The Wasserstein-1 distance is then combined with the loss function on the new data for incremental training, and the current iteration parameters are guided to move closer to the historical parameters. The degree of closeness is adjusted according to the distribution difference, which can achieve robust adaptation to scenarios where the distributions of new and old data are inconsistent. The calculation formula is as follows: , , in, for The Wasserstein-1 distance between the old and new data distributions at time point represents the difference between the old and new data distributions. The Wasserstein-1 distance calculation function; This represents the input data; Distribution of old data; For the new data distribution; This is the set of trainable parameters for the updated LSTM model; This is the set of trainable parameters for the old LSTM model; This is the set of trainable parameters for the LSTM model in the current iteration during incremental training. Set the learning rate for incremental training. ; For gradient operators; The loss function for incremental training on new data; The alignment intensity coefficient is used to adjust for the impact of distribution differences on parameter updates. ; S5.3, Evaluate the incremental training effect: After incremental training is completed, the performance of the irrigation district agricultural irrigation water consumption estimation model is evaluated using an independent test set. If the prediction accuracy recovers to the preset standard and the performance is balanced on the old and new data, then update the parameters of the irrigation district agricultural irrigation water consumption estimation model; otherwise, adjust the alignment strength coefficient and retrain.

[0044] Example 2 An irrigation district water consumption estimation system based on adaptive dynamic feature fusion includes a module for executing processing instructions for each step in an irrigation district water consumption estimation method based on adaptive dynamic feature fusion. Includes the following modules: Sensing and data acquisition module: Collects water consumption data of the irrigation area through equipment deployed in various levels of canals in the irrigation area, and transmits the collected data to a central database for storage via the LoRaWAN protocol; Data preprocessing module: Reads water consumption data collected from the central database, performs data cleaning, time alignment and normalization, and constructs a high-quality training dataset; Feature fusion engine module: includes feature enhancement module, dynamic lag embedding module and periodic gating fusion module; real-time meteorological covariates are injected into historical water consumption information to construct dynamic lag embedding matrix, and then fused with periodic features to obtain historical fused features; Irrigation water consumption estimation module for irrigation districts: The preprocessed data and historical fusion features are input into the irrigation district agricultural irrigation water consumption estimation module for training until the model meets the preset conditions and then training stops. Incremental learning module: Incrementally trains the agricultural irrigation water consumption estimation module in the irrigation area after training; Application Decision Module: Input the new real-time collected irrigation district quantity data and meteorological data into the irrigation district agricultural irrigation water consumption estimation module after incremental training to estimate the irrigation district agricultural irrigation water consumption.

[0045] Example 3 like Figure 2 As shown, to verify the advantages of the dynamic lag embedding matrix in capturing the dynamic time-delay relationship between meteorological events and irrigation responses, the prediction effects of three methods—fixed lag embedding, sliding window method, and the dynamic lag embedding of this invention—on actual water consumption were compared. The horizontal axis represents time (days), and the vertical axis represents irrigation water consumption (cubic meters). The two key abrupt change points, drought and rainstorm events, are marked with gray dashed lines. The experimental results clearly show that when drought events (day 12) and rainstorm events (day 47) occur, both fixed lag embedding and sliding window method show significant prediction lags. The blue dashed line and green dotted line deviate from the actual black curve at the event points, while the dynamic lag embedding (red solid line) almost coincides with the actual water consumption curve. This proves that the method of determining the optimal lag step size through time-varying mutual information in this invention can effectively capture the dynamic time-delay relationship between meteorological events and irrigation responses, avoiding the distortion problem in capturing abrupt change features caused by the fixed lag step size in conventional methods.

[0046] Example 4 like Figures 3 to 5 As shown, to evaluate the overall performance difference between this invention and mainstream prediction methods in terms of irrigation water consumption estimation accuracy, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination are compared. Three key indicators visually demonstrate the advantages of this invention over traditional long short-term memory network models, Prophet time series prediction models, autoregressive composite moving average models, and support vector regression models. Figures 3 to 5As can be seen, the blue bars representing the present invention are significantly superior to other methods in all three subgraphs. They show the lowest values ​​for the two negative indicators, root mean square error and mean absolute error, while reaching the highest value for the positive indicator, coefficient of determination. This verifies the effectiveness of the dynamic hysteresis embedding matrix and periodic gating fusion mechanism, indicating that the present invention can more accurately capture the complex patterns of irrigation water volume changes and provide a more reliable basis for irrigation decisions.

[0047] Example 5 like Figure 6 As shown, to evaluate the responsiveness of this invention to sudden irrigation water events, a real-world scenario involving two sudden water usage peaks was simulated. The prediction performance of this invention was compared with that of the traditional Long Short-Term Memory (LSTM) network model and the Autoregressive Integrated Moving Average (ARIMA) model. In the figure, the black true curve shows obvious peaks at the 10-day and 20-day positions, representing sudden irrigation water events. The blue prediction line (this invention) is closest to the true value at the point of change, with the smallest lag effect. The green prediction line (traditional LSTM) shows obvious lag and underestimation of magnitude. The red prediction line (ARIMA) shows the most severe prediction bias, highlighting the design value of the gradient-enhanced forget gate in this invention. By injecting mutation information through a normalized gradient factor, the model can quickly respond to extreme irrigation events, solving the key defect of conventional models being slow to respond to sudden events.

[0048] Example 6 like Figure 7 As shown, to analyze the differences in optimization efficiency of different methods during model training, a logarithmic coordinate system is used to display the relationship between training epochs and loss values. The blue curve (the present invention) shows the steepest initial descent trend and the lowest final convergence value. Compared with the green curve of the traditional Long Short-Term Memory Network (traditional LSTM) and the red curve of the Prophet time series prediction model, the present invention reaches a stable convergence state in about 40 epochs, while other methods require more training epochs and have higher final loss values. This demonstrates the dual advantages of gradient-oriented pruning optimization and memory unit sparsity regularization. It not only solves the optimization oscillation problem caused by the coexistence of small gradient stable segments and large gradient abrupt segments in irrigation data, but also improves the model's generalization ability through structured sparsity penalty, significantly accelerating the training process and improving the final model quality.

[0049] Example 7 like Figure 8As shown, to analyze the error distribution characteristics of various prediction methods under different seasonal conditions, box plots are used to show the dispersion of prediction errors during peak season (peak irrigation season) and off-season through interquartile range and outlier distribution. This invention exhibits the smallest box height and the most concentrated outlier distribution in both seasons, indicating that the prediction results are the most stable and reliable. In particular, during the peak season, the box position of this invention is significantly lower than that of other methods, verifying the effectiveness of the seasonality importance factor in the loss function design. This mechanism significantly improves the prediction accuracy of key agricultural periods by strengthening the peak season error penalty. In contrast, traditional long short-term memory networks and Prophet models show a large number of high-level outliers during the peak season, reflecting the limitations of the conventional mean square error equalization processing mode in dealing with seasonal differences.

[0050] Example 8 like Figure 9 As shown, to test the robustness of the incremental training mechanism to data distribution drift, the horizontal axis of the line graph represents time (days), and the vertical axis represents the moving average absolute error (7-day window). The gray dashed line marks the start of the dry season (day 90). The fixed window fine-tuning method (blue line) shows a sharp increase in error and sustained high fluctuations after the seasonal transition, indicating its inability to adapt to changes in water use patterns caused by drought. In contrast, the incremental training method of this invention (red line) only shows a brief increase in error at the beginning of the seasonal transition before quickly falling back. This demonstrates that the incremental online temporal alignment mechanism designed in this invention dynamically adjusts the parameter update direction by calculating the difference between the old and new data distributions, enabling the model to retain historical knowledge while quickly adapting to the new distribution, effectively solving the data distribution drift problem caused by irrigation infrastructure renovation or sudden climate changes.

[0051] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. An irrigation area water consumption estimation method based on adaptive dynamic feature fusion, characterized in that, Includes the following steps: S1. By deploying smart water meters, flow meters and remote sensing monitoring equipment in the irrigation area at all levels, the water consumption data of the irrigation area is collected in real time, and then the collected data is transmitted to the central database for storage through the Internet of Things terminal. S2. Read the collected water consumption data from the central database and perform data cleaning, time alignment and normalization. S3. Construct a dynamic lag embedding matrix based on historical water consumption information, and then fuse it with periodic features to obtain historical fused features; S3 is as follows: S3.1 Constructing a dynamic lag embedding matrix for historical water consumption data: Based on the time-varying mutual information between irrigation district water consumption data and meteorological factors, a dynamic lag embedding matrix is ​​constructed. First, the time-varying mutual information between historical water consumption data and meteorological factors is calculated to determine the optimal lag step size for each time point. Then, according to the determined optimal lag step size, water consumption observations for each irrigation district at different lag time points are extracted from historical data. The water consumption observations are arranged in order of irrigation district and lag step size to form a dynamic lag embedding matrix. Each row of the matrix corresponds to the water consumption of an irrigation district at different lag time points, and each column corresponds to the water consumption of all irrigation districts at a specific lag time point. The dynamic lag embedding matrix of historical water consumption data is represented as follows: , wherein, is a dynamic lag embedding matrix; is the water use observation value of irrigation district 1 at time ; is the water use observation value of irrigation district 1 at time ; is the water use observation value of irrigation district Na at time ; is the water use observation value of irrigation district Na at time ; is the current time index; is the maximum lag step size; is the minimum lag step size; is the maximum lag order; is the number of irrigation districts; S3.2, Perform periodic feature fusion of irrigation district water consumption data: In the process of constructing periodic features, time embedding features in two dimensions—daily and monthly—are extracted based on the current time index. The daily period embedding represents the phased changes in daily irrigation activities, while the monthly period embedding represents the macro-adjustment trend of planting structure. Then, the two periodic features are input into a weight transformation function to calculate the periodic gating vector. The periodic gating vector is used to control the fusion ratio between historical water consumption features and periodic features, thereby generating a periodically guided feature vector. The fused feature vector is calculated using the following formula: , , , , in, This represents the daily periodic eigenvector; This represents the monthly cycle feature vector; For scalar splicing; It is a periodic gated vector; Use the Sigmoid activation function; This is element-wise multiplication; The characteristics of historical integration after fusion; This is the weight matrix for the daily cycle characteristics; This is the weight matrix for the monthly cycle characteristics; The weight matrix for the periodic feature transformation; This indicates vector concatenation; For the dynamic hysteresis embedding matrix in time The input vector; S4. Construct an agricultural irrigation water consumption estimation model for irrigation districts based on a deep LSTM model. Input the preprocessed data and historical fusion features into the model for training until the model meets the preset conditions and then stop training to obtain a trained agricultural irrigation water consumption estimation model for irrigation districts. S5. Incremental training is performed on the trained irrigation water consumption estimation model for the irrigation district. S6. Input the newly collected irrigation water consumption data and meteorological data into the incrementally trained model to estimate the agricultural irrigation water consumption in the irrigation area.

2. The irrigation district water consumption estimation method based on adaptive dynamic feature fusion according to claim 1, characterized in that, S4 is as follows: A deep LSTM model is constructed to estimate agricultural irrigation water consumption in irrigation districts. The model is a deep network structure containing multiple LSTM units. Each LSTM unit extracts temporal features at different time scales. The first LSTM unit receives the fused feature vector as input and gradually abstracts higher-order temporal dependencies through stacked hidden layers. The last LSTM unit is connected to the multi-scale feature pyramid output layer to fuse the daily and weekly hidden states into the final predicted value. The preprocessed data and the fused historical features are input into the model for training. First, the trainable parameters of the model are initialized, the weight matrix and bias term of the LSTM unit are initialized with a normal distribution, and the initial variance and forget gate bias term are set. Then, the model is trained by improving the forward propagation process of the deep LSTM model, constructing a multi-scale feature pyramid output, calculating the loss function, performing gradient-oriented pruning optimization, and performing memory cell sparsity regularization. Mini-batch gradient descent is used to train the model for each iteration. Each batch samples time-series segments from the preprocessed data and inputs them into the deep LSTM model. After forward propagation generates predicted values, the parameter update amount is calculated using the calculated loss function and the gradient of the loss function with respect to the trainable parameters. During backpropagation, the model parameters are updated in conjunction with the gradient pruning strategy. The learning rate of the deep LSTM model adopts an exponential decay strategy until the loss function does not decrease significantly in consecutive iterations or the difference between the training error and the validation error is lower than a preset threshold. Once the model meets the preset conditions, training stops, resulting in a well-trained irrigation water estimation model for the irrigation district.

3. The irrigation district water consumption estimation method based on adaptive dynamic feature fusion according to claim 2, characterized in that, The specific steps for improving the forward propagation process of the deep LSTM model are as follows: The water consumption input at the current time step is jointly projected onto the hidden state at the previous time step, and a normalized gradient factor is constructed by combining the difference between the water consumption at the current time step and the previous time step. The normalized gradient factor is injected into the forget gate in a weighted form to obtain the improved forget gate. At the same time, during the state update process, the previous state and the candidate state determine the current memory cell state through the gate control mechanism.

4. The irrigation district water consumption estimation method based on adaptive dynamic feature fusion according to claim 3, characterized in that, The specific steps for constructing the multi-scale feature pyramid output are as follows: First, the input at the current time is encoded using a standard LSTM unit to obtain the daily granular hidden state. Then, max pooling is performed on the daily granular hidden state every 7 time steps to extract the most salient features within a week, forming the weekly granular hidden state. Finally, in the output stage, the current daily granular hidden state and its corresponding weekly granular hidden state are weighted and fused to output the final water consumption prediction result.

5. The irrigation district water consumption estimation method based on adaptive dynamic feature fusion according to claim 4, characterized in that, The calculation process of the loss function is as follows: Based on the deviation between the predicted and actual results, an error term is constructed for each time step. A predefined error threshold is used to determine whether the error needs to be penalized. If the error is greater than the error threshold, the seasonal importance factor is used to enhance its weight in the loss function. Otherwise, the basic error weight is maintained. Finally, the loss terms of all time steps are weighted and averaged, and then combined with a regularization term to obtain the overall loss.

6. The irrigation district water consumption estimation method based on adaptive dynamic feature fusion according to claim 5, characterized in that, gradient The specific process for targeted cropping optimization is as follows: During the model parameter update phase, the gradient vector at the current time step is first recorded, and the average direction is calculated based on the gradient values ​​of several historical steps. If the current gradient value exceeds the clipping threshold and the direction is consistent with the historical gradient direction, the current gradient size is scaled according to the historical average direction, and the gradient with the out-of-limit value is clipped in a directed manner.

7. The irrigation district water consumption estimation method based on adaptive dynamic feature fusion according to claim 6, characterized in that, The specific process of sparse regularization of memory units is as follows: The LSTM memory states are divided into several groups. The L2 norm of each group of memory cells is calculated, and a weight term based on the group size is applied to form a structured regularization penalty term. The calculation formula is as follows: , in, For regularization terms; Number of groups for memory units; Grouped index; Represents the norm symbol; For the first Group memory units; Indicates the first The number of group memory units; The sparsity strength is denoted by .

8. The irrigation district water consumption estimation method based on adaptive dynamic feature fusion according to claim 7, characterized in that, S5 is detailed below: S5.1 Define the incremental training trigger conditions: Incremental training is automatically triggered when the distribution difference between the newly added water consumption data and the historical data exceeds the threshold, or when the prediction error of the irrigation district agricultural irrigation water consumption estimation model on the latest data continues to be higher than the acceptable threshold. S5.2 Incremental Online Timing Alignment: During incremental training, an incremental online temporal alignment method is adopted. When using new data for incremental model training, the Wasserstein-1 distance between the distributions of the new and old sample data is first calculated. The Wasserstein-1 distance is then combined with the loss function on the new data for incremental training, and the current iteration parameters are guided to move closer to the historical parameters. The degree of closeness is adjusted according to the distribution differences. The calculation formula is as follows: , , in, for Wasserstein-1 distance between the distributions of new and old data at time points; The Wasserstein-1 distance calculation function; This represents the input data; Distribution of old data; For the new data distribution; This is the set of trainable parameters for the updated LSTM model; This is the set of trainable parameters for the old LSTM model; This is the set of trainable parameters for the LSTM model in the current iteration during incremental training. Incremental training learning rate; For gradient operators; The loss function for incremental training on new data; Alignment strength coefficient; S5.3, Evaluate the incremental training effect: After incremental training is completed, the performance of the irrigation district agricultural irrigation water consumption estimation model is evaluated using an independent test set. If the prediction accuracy recovers to the preset standard and the performance is balanced on the old and new data, then update the parameters of the irrigation district agricultural irrigation water consumption estimation model; otherwise, adjust the alignment strength coefficient and retrain.

9. A system for estimating irrigation water consumption based on adaptive dynamic feature fusion, comprising executing the method for estimating irrigation water consumption based on adaptive dynamic feature fusion as described in any one of claims 1-8, characterized in that, Includes the following modules: Sensing and data acquisition module: Collects water consumption data of the irrigation area through equipment deployed in various levels of canals in the irrigation area, and transmits the collected data to a central database for storage via the LoRaWAN protocol; Data preprocessing module: Reads water consumption data collected from the central database, performs data cleaning, time alignment and normalization, and constructs a high-quality training dataset; Feature fusion engine module: It includes a feature enhancement module, a dynamic hysteresis embedding module, and a periodic gating fusion module; Real-time meteorological covariates are injected into historical water consumption information to construct a dynamic lag embedding matrix, which is then fused with periodic features to obtain historical fused features. Irrigation water consumption estimation module for irrigation districts: The preprocessed data and historical fusion features are input into the irrigation district agricultural irrigation water consumption estimation module for training until the model meets the preset conditions and then training stops. Incremental learning module: Incrementally trains the agricultural irrigation water consumption estimation module in the irrigation area after training; Application Decision Module: Input the newly collected irrigation water consumption data and meteorological data in real time into the irrigation area agricultural irrigation water consumption estimation module after incremental training to estimate the irrigation area agricultural irrigation water consumption.

Citation Information

Patent Citations

  • Method for monitoring scale irrigation water consumption of irrigation area based on physical constraint machine learning

    CN120611626A