Irrigation area water volume prediction and control system based on artificial intelligence

Abstract

The invention belongs to the technical field of intelligent water conservancy, and particularly relates to an irrigation area water volume prediction and control system based on artificial intelligence. Firstly, a heterogeneous data acquisition module synchronously acquires multi-dimensional environment data from different sensors, and accurate synchronization of the data is ensured by adopting LoRa, NB-IoT and 4G protocols; secondly, the collaborative restoration module detects and restores hardware drift and signal conflicts in sensor data in combination with a space-time constrained graph attention network, and space-time calibration is carried out through causal reasoning; and finally, the increment prediction module generates adaptive parameters based on the element network, outputs a water volume prediction result through the Bayesian neural network, and generates a regulation and control instruction in combination with the confidence coefficient parameter. According to the invention, different climate conditions and emergencies can be accurately coped with, the intelligent level of water resource management is improved, and the precision of water volume prediction and the regulation and control reliability of irrigation area water resources are improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent water conservancy, and particularly relates to an irrigation area water quantity prediction and control system based on artificial intelligence. BACKGROUND

[0002] Irrigation area water quantity prediction and control is a core link of agricultural water resource management, involves meteorology, hydrology, agronomy and automatic control and other multidisciplinary cross technologies, and traditional irrigation area management mainly relies on historical hydrological data, crop water demand models and artificial experience for water quantity scheduling. The decision-making process needs to comprehensively consider multiple factors such as rainfall, evaporation, soil moisture, crop growth stage and the like. With the development of sensor technology, Internet of Things and data science, modern irrigation area systems have gradually realized automatic collection and remote monitoring of hydrological data, laying a foundation for fine irrigation.

[0003] However, the current irrigation area water quantity prediction and control system relies on multiple types of sensors, which can easily cause time delay differences in data collection, so that the collected environmental data cannot be synchronized, thereby affecting the real-time performance and accuracy of water quantity prediction and control. Secondly, in the process of long-time operation, the sensors are prone to hardware drift or signal conflict, affecting the accuracy of the data. The existing technology has limited processing capacity for these problems and cannot provide an automatic repair mechanism, so that the system cannot correct the data error in time, thereby affecting the accuracy of water quantity prediction and the reliability of decision-making. Finally, traditional water quantity prediction technology mainly relies on historical hydrological data and static crop water demand models, and cannot quickly adapt to environmental changes or emergencies. In particular, in the face of different seasons, climate fluctuations or sudden weather events, the existing prediction model cannot be adjusted or updated in time, resulting in a decrease in prediction accuracy and the inability to provide effective decision-making basis for irrigation area management.

[0004] Therefore, the application provides an irrigation area water quantity prediction and control system based on artificial intelligence. SUMMARY

[0005] In order to solve the technical problems mentioned in the background art, the purpose of the present application is to provide an irrigation area water quantity prediction and control system based on artificial intelligence.

[0006] In order to achieve the above-mentioned purpose, the application adopts the following technical scheme:

[0007] An irrigation area water quantity prediction and control system based on artificial intelligence comprises:

[0008] M1: a heterogeneous data collection module, deploying sensors in the irrigation area and synchronously collecting multi-dimensional environmental data through a heterogeneous communication protocol;

[0009] M2: A collaborative repair module, which constructs a sensor physical topology graph based on the multi-dimensional environment data, detects hardware drift and signal conflict using a spatio-temporal constraint graph attention network, and outputs spatio-temporal calibration data in combination with a causal reasoning loss function;

[0010] M3: An incremental prediction module, which receives the spatio-temporal calibration data, generates scene adaptive parameters based on a meta-network, fuses historical data distribution and real-time features through incremental learning, and outputs water quantity prediction results and confidence parameters using a Bayesian neural network;

[0011] M4: A reinforcement control module, which generates control instructions based on the water quantity prediction results and confidence parameters, and updates prior parameters of the meta-network in a reverse gradient manner based on feedback signals of instruction execution.

[0012] Further, the sensors include a rainfall sensor, a temperature and humidity sensor, a soil moisture content sensor, a water level sensor, and a flow sensor;

[0013] Wherein, the rainfall sensor data symbol is represented as R(u, j); the temperature and humidity sensor data symbol is represented as T(u, j); the soil moisture content sensor data symbol is represented as S(u, j); the water level sensor data symbol is represented as W(u, j); and the flow sensor data symbol is represented as Q(u, j); wherein u represents a time round of data collection; and j represents data collected by the corresponding sensor.

[0014] Further, the heterogeneous communication protocols include LoRa protocol, NB-IoT protocol, and 4G protocol, the LoRa protocol is suitable for the rainfall sensor and the soil moisture content sensor, the communication time slot length is 10 seconds; the NB-IoT protocol is suitable for the temperature and humidity sensor and the flow sensor, the communication time slot length is 8 seconds; and the 4G protocol is suitable for the water level sensor, the communication time slot length is 6 seconds.

[0015] Further, the sensor physical topology graph is constructed by the spatial positions of the sensors, so as to reflect the relationship between the sensors, and the geographic coordinates of the sensor p and the sensor q are defined as (x p ,y p ) and (x q ,y q ) respectively, and the distance between the sensor p and the sensor q is D pq ;

[0016] In the spatio-temporal constraint graph attention network, the relationship between the sensors is represented by a graph structure, and the data transmission between the sensors is weighted using an attention mechanism, and the data collected by each sensor p at time t is x p,t, represents the data feature of sensor p at time t, the attention weight a in the graph attention network pq reflects the influence degree of sensor p on sensor q, and the calculation formula is:

[0017]

[0018] wherein, exp(*) is an exponential function; LR(*) is a LeakyReLU activation function; b is a weight vector, representing a learning parameter related to the data of each sensor; b T is the transpose of the weight vector b, indicating the conversion of its column vector to a row vector; W is a learning weight matrix, used to map the sensor data x p and x q to a new feature space; Wx p represents the linear transformation result of the feature of sensor p; x p and x q respectively represent the feature vectors of sensor p and sensor q at time t; ‖ is a concatenation operation; is a neighbor set of sensor p, representing a set of sensors directly connected to sensor p in the graph; Wx r is the linear transformation result of the feature of neighbor sensor r;

[0019] The linear transformation result Wx r is calculated through the weight matrix W and the input data x r of sensor r, and the formula is Wx r = W·x r , wherein the weight matrix W represents the linear mapping of the feature data of sensor p in sensor r;

[0020] The attention weight a pq is used to weight and fuse the data of sensor p, and the formula is:

[0021]

[0022] wherein, is the data representation of sensor p after weighted fusion;

[0023] Based on the data after weighted fusion, the sensor prediction value of the graph attention network is generated

[0024]

[0025] wherein, W1 is the first layer weight matrix, used to map the input data Linear transformation is performed; W2 is a second layer weight matrix, used to process data after activation function processing; b1 is a first layer bias term, used for data from the input layer to the hidden layer; b2 is a second layer bias term, used for data from the hidden layer to the final output layer; ReLU(*) is a ReLU activation function, representing ReLU(x) = max(0, x), used to introduce a nonlinear transformation, and x represents input data passed to the ReLU activation function.

[0026] Further, the sensor actual measurement value y t The actual observation data provided by the sensor is directly obtained by the sensor, and the sensor prediction value is compared with the sensor actual measurement value y t , and the error loss is calculated as follows:

[0027]

[0028] Wherein, L e is an error loss function of hardware drift and signal conflict, used to measure the difference between the sensor actual measurement value and the sensor prediction value; N is the length of the time series, representing the total length of the observation data;

[0029] The causal inference loss function calibrates the sensor data by processing the causal relationship of the spatio-temporal data using a minimum loss function, and the formula is:

[0030]

[0031] Wherein, L ca is a loss function of causal inference, representing the difference between the prediction value and the actual observation value; θ is a parameter of the causal inference model, representing the learning result of the model; f(x t , θ) is a prediction value generated by the causal inference model based on input data x t and parameter θ; is an L2 norm, representing the squared error between the prediction value and the actual value;

[0032] Finally, the spatio-temporal calibrated sensor output is the spatio-temporal calibration data X', and X' = X'1, X'2, …, X'N. p .

[0033] Further, the meta-network generates adaptive parameters suitable for different tasks by learning from multiple tasks:

[0034]

[0035] Wherein, is an optimal parameter, representing a parameter that minimizes the loss function ; representing finding the parameters that minimize the loss function T is a task, representing a specific data processing task; E T～p(T) representing taking the expectation of the task distribution p(T); is the loss function of the task T, used to measure the prediction error of the meta-network model under the task; is the output of the meta-network model with given meta-network parameters , the prediction result calculated based on the input data X';

[0036] The incremental learning updates the parameters of the meta-network model by combining historical data and real-time data, and the incremental learning formula is:

[0037]

[0038] wherein, represents the meta-network model parameters at time t+1; represents the meta-network model parameters at time t; α is the learning rate, used to control the amplitude of parameter update; is the gradient;

[0039] The updated meta-network model uses the Bayesian neural network for water quantity prediction to provide a confidence level for each prediction value, reflecting the uncertainty of the prediction result:

[0040]

[0041] wherein, is the water quantity prediction result y output by the Bayesian neural network when given input data X and meta-network model parameters ; represents different values of the Bayesian neural network model parameters; is the Bayesian posterior distribution; is the integral of the Bayesian neural network model parameters ;

[0042] The confidence parameter σ is calculated based on the Bayesian posterior distribution and the integral , and the formula is

[0043] Finally, the water quantity prediction result y and the confidence parameter σ are output.

[0044] Further, the control instruction u i is generated based on the water quantity prediction result y and the confidence parameter σ, and the formula is as follows:

[0045] u i =c1·y+c2·σ+k

[0046] wherein c1 and c2 are weight parameters, which are used to control the influence of the water quantity prediction result and the confidence parameter on the control instruction respectively; k is a bias term, which is used to adjust the reference value of the instruction;

[0047] According to the execution feedback of the control instruction u i , the feedback signal is defined as r i , and the gradient reverse update formula of the loss function is:

[0048]

[0049] wherein φ t+1 is the updated meta-network model prior parameter; φ t is the prior parameter of the current meta-network model; z is a learning rate, which is used to control the update amplitude; is the gradient of the control instruction u i on the meta-network model prior parameter φ t .

[0050] Compared with the prior art, the present application has the following advantages:

[0051] 1. The present application uses three communication protocols of LoRa, NB-IoT and 4G through the heterogeneous data acquisition module, and synchronously acquires multi-dimensional data such as rainfall, temperature and humidity, and soil moisture content, so that the system can ensure that the data of different types of sensors are accurately synchronized in the same time window, eliminating the data delay and time misalignment problems caused by the non-uniformity of sensor protocols, thereby ensuring the efficient accuracy of the water quantity prediction in the irrigation area.

[0052] 2. The present application uses the cooperative repair module combined with the spatiotemporal constraint graph attention network, so that the system can automatically detect and repair the data errors caused by hardware drift and signal conflict. This technology can calibrate the sensor data in real time, ensure that the output data of each sensor is accurately repaired in physics and time, and greatly improve the accuracy of water quantity prediction and the reliability of water resource regulation in the irrigation area.

[0053] 3. The present application uses the incremental prediction module combined with historical data and real-time data, generates scene adaptive parameters through the meta-network, and adjusts the prediction model parameters in time through the incremental learning method, so that the system can adapt to changes in different environments and climate conditions. This mechanism enables the system to quickly update the prediction model when dealing with seasonal changes, climate mutations and other factors, and provides more accurate water quantity prediction results. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort based on these drawings.

[0055] Fig. 1 The system workflow schematic diagram of the present application;

[0056] Fig. 2 The multi-dimensional environment data acquisition process schematic diagram of the present application;

[0057] Fig. 3 The meta-network model parameter updating process schematic diagram of the present application. DETAILED DESCRIPTION

[0058] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort are within the protection scope of the present application.

[0059] To achieve the above object, the present application is implemented by the following technical solutions. The present application provides a water quantity prediction and control system for irrigation area based on artificial intelligence, as shown in the figure, the system comprises: Figs. 1-3

[0060] M1: Heterogeneous data acquisition module, deploy sensors in the irrigation area, and synchronously collect multi-dimensional environment data through heterogeneous communication protocol.

[0061] The sensors comprise rainfall sensor, temperature and humidity sensor, soil moisture content sensor, water level sensor and flow sensor;

[0062] The rainfall sensor is used to measure the precipitation, and the data symbol is represented as R(u,j), and the physical acquisition range is 0-500mm; the temperature and humidity sensor is used to measure the environment temperature and humidity, and the data symbol is represented as T(u,j), and the physical acquisition range is-40℃ to 60℃; the soil moisture content sensor is used to measure the water content of soil, and the data symbol is represented as S(u,j), and the physical acquisition range is 0%-100%; the water level sensor is used to measure the water level of the water source in the irrigation area, and the data symbol is represented as W(u,j), and the physical acquisition range is 0-10m; the flow sensor is used to measure the flow of irrigation water flow, and the data symbol is represented as Q(u,j), and the physical acquisition range is 0-50m 3 ​ / s;

[0063] The above data is arranged in a table as follows:

[0064]

[0065]

[0066] wherein a is the sensor number, representing the category number of each sensor; u represents the time round of data collection; j represents the data collected by the corresponding sensor;

[0067] The heterogeneous communication protocol includes LoRa protocol, NB-IoT protocol and 4G protocol, each type of sensor is classified according to the communication protocol used, and the corresponding communication time slot is allocated, and the correspondence between different sensors and communication protocols is as follows:

[0068] The LoRa protocol is suitable for the rainfall sensor and the soil moisture content sensor, and the communication time slot length is 10 seconds; the NB-IoT protocol is suitable for the temperature and humidity sensor and the flow sensor, and the communication time slot length is 8 seconds; the 4G protocol is suitable for the water level sensor, and the communication time slot length is 6 seconds.

[0069] M2: collaborative repair module, based on the multi-dimensional environment data, constructing a sensor physical topology graph, using a spatio-temporal constraint graph attention network to detect hardware drift and signal conflict, and combining a causal reasoning loss function to output spatio-temporal calibration data.

[0070] The sensor physical topology graph is constructed by the spatial position of the sensor, which reflects the relationship between sensors. In this embodiment, the geographical coordinates of the sensor p and the sensor q are defined as (x p ,y p ) and (x q ,y q ), respectively, and the distance D pq between the sensor p and the sensor q is calculated as follows:

[0071] In the spatio-temporal constraint graph attention network, the relationship between the sensors is represented by a graph structure, and the data transmission between the sensors is weighted by using an attention mechanism. The data collected by each sensor p at time t is x p,t , which represents the data features of sensor p at time t. The attention weight a pq in the graph attention network reflects the influence degree of sensor p on sensor q, and the calculation formula is as follows:

[0072]

[0073] where exp(*) is an exponential function, representing e x ; LR(*) is a LeakyReLU activation function; b is a weight vector, representing learning parameters related to data of each sensor; b T is a transpose of the weight vector b, representing converting its column vector into a row vector; W is a learning weight matrix, used to map sensor data x p and x q to a new feature space, specifically, W is an m x n matrix, where m is the dimension of the output space, n is the dimension of the input space, and each column represents a feature of the input data; Wx p represents a linear transformation result of the feature of sensor p, representing the new feature space; x p and x q represent the feature vectors of sensor p and sensor q at time t, respectively; ‖ is a concatenation operation, representing concatenating two vectors into a longer vector; is a neighbor set of sensor p, representing a set of sensors directly connected to sensor p in the graph; Wx r is a linear transformation result of the feature of neighbor sensor r;

[0074] The linear transformation result Wx r is calculated by the weight matrix W and the input data xr of sensor r, and the formula is Wx r = W·x r , where the weight matrix W represents a linear mapping of the feature data of sensor p in sensor r;

[0075] The LeakyReLU activation function LR(*) is used to introduce nonlinear features for sensor data to prevent dead neuron problems, and is defined as LR(*) = max(0.01·*, *).

[0076] The attention weight a pq is used to weight and fuse the data of sensor p, and the formula is represented as:

[0077]

[0078] where, is a representation of the data of sensor p after weighted fusion;

[0079] Based on the data after weighted fusion a sensor prediction value of the graph attention network is generated

[0080]

[0081] where W1 is a first layer weight matrix, used to map the input data linear transformation; W2 is a second layer weight matrix, used to process data after activation function processing; b1 is a first layer bias term, used for data from the input layer to the hidden layer; b2 is a second layer bias term, used for data from the hidden layer to the final output layer; ReLU(*) is a ReLU activation function, representing ReLU(x) = max(0, x), used to introduce a nonlinear transformation, and x represents input data passed to the ReLU activation function;

[0082] sensor actual measurement value y t actual observation data provided by the sensor, directly obtained by the sensor, and the sensor prediction value is compared, and the error loss is calculated as follows: t

[0083]

[0084] wherein L e is an error loss function of hardware drift and signal conflict, used to measure the difference between the sensor actual measurement value and the sensor prediction value; N is the length of the time series, representing the total length of the observation data;

[0085] The causal inference loss function calibrates the sensor data by processing the causal relationship of the spatio-temporal data using a minimum loss function, and the causal inference model adjusts the parameters by inputting data x t and the sensor actual measurement value y t , and the formula is:

[0086]

[0087] wherein L ca is a loss function of causal inference, representing the difference between the prediction value and the actual observation value; θ is a parameter of the causal inference model, representing the learning result of the model; f(x t , θ) is a prediction value generated by the causal inference model based on input data x t and parameter θ; is an L2 norm, representing the squared error between the prediction value and the actual value;

[0088] Finally, the spatio-temporal calibrated sensor output is the spatio-temporal calibrated data X', and X' = X'1, X'2, …, X'N. p .

[0089] M3: Incremental prediction module, receiving the spatio-temporal calibrated data, generating scene adaptive parameters based on the meta-network, fusing historical data distribution and real-time features through incremental learning, and outputting water quantity prediction results and confidence parameters using a Bayesian neural network.

[0090] ​The meta-network generates adaptive parameters suitable for different tasks by learning from multiple tasks:

[0091]

[0092] wherein, is an optimal parameter, indicating a parameter that causes the loss function to be minimized; indicates a parameter that minimizes the loss function; T is a task, indicating a specific data processing task, such as water volume prediction tasks in different seasons, irrigation area analysis in different regions, etc. T～p(T) represents the expectation of the task distribution p(T) to ensure that the meta-network model can generalize to multiple tasks; is the loss function of task T, used to measure the prediction error of the meta-network model under this task; is the output of the meta-network model with given meta-network parameters , the prediction result calculated based on the input data X';

[0093] The incremental learning updates the parameters of the meta-network model by combining historical data and real-time data, ensuring that the meta-network model can adapt to new data features over time. The incremental learning formula is:

[0094]

[0095] wherein, represents the meta-network model parameter at time t+1; represents the meta-network model parameter at time t; α is the learning rate, used to control the magnitude of parameter update; is the gradient, indicating how to update the meta-network model according to the current data point χ t and the current parameter ;

[0096] The updated meta-network model uses the Bayesian neural network for water volume prediction to provide confidence for each prediction value, reflecting the uncertainty of the prediction result:

[0097]

[0098] wherein, is the water volume prediction result y output by the Bayesian neural network given the input data X and the meta-network model parameter ; represents different values of the Bayesian neural network model parameter, indicating the randomness of the Bayesian neural network model parameter; is the Bayesian posterior distribution, indicating the uncertainty of the Bayesian neural network model parameter; To integrate the Bayesian neural network model parameters To obtain the final prediction result by weighted average of all possible parameter values;

[0099] The confidence parameter σ is calculated based on the Bayesian posterior distribution And the integral The formula is

[0100] Finally, the water volume prediction result y and the confidence parameter σ are output.

[0101] M4: Reinforcement control module, based on the water volume prediction result and the confidence parameter to generate control instructions, and based on the feedback signal of instruction execution, the gradient of the prior parameter of the meta-network is updated in reverse.

[0102] Based on the water volume prediction result y and the confidence parameter σ to generate the control instruction u i The formula is as follows:

[0103] u i =c1·y+c2·σ+k

[0104] Wherein, c1 and c2 are weight parameters, which are used to control the influence of water volume prediction result and confidence parameter on control instruction in turn; k is a bias term, which is used to adjust the baseline value of the instruction;

[0105] In this embodiment, the value of the weight parameter c1 is 0.8; the value of the weight parameter c2 is 0.2; k=5 is a regulation constant, which is obtained based on actual experimental data and system optimization adjustment, to ensure the stability and optimization effect in model operation;

[0106] According to the execution feedback of the control instruction u i The feedback signal is defined as r i And the gradient reverse update formula of the loss function is:

[0107]

[0108] Wherein, φ t+1 Is the updated prior parameter of the meta-network model; φ t Is the prior parameter of the current meta-network model; z is the learning rate, which is used to control the update amplitude; Is the gradient of the control instruction u i To the prior parameter φ t Of the meta-network model;

[0109] In this embodiment, the value of the learning rate z is 0.01;

[0110] The meta-network model parameter is a model parameter obtained by the causal inference model in a training process, and is a model parameter for a single task or a specific field, which is optimized by a training task; the meta-network model prior parameter is an initial parameter given by a loss function when the meta-network model faces different tasks or environments, which is shared between multiple tasks and adjusted by the meta-network model to adapt to different task environments;

[0111] The embodiment provides an example:

[0112] Suppose that a water quantity prediction and control system in an irrigation area needs to generate a control instruction u i to control the water flow, the system predicts water quantity demands of 30 mm, 40 mm and 50 mm in the next three days, and the corresponding current confidence is 0.85;

[0113] According to the input of the system, the specific control formula is:

[0114] u i = 0.8 * 40 + 0.2 * 0.85 + 5 = 32 + 0.17 + 5 = 37.17

[0115] That is, the generated water quantity control instruction is 37.17;

[0116] Suppose that the actual feedback signal r i of the system is 0.9, indicating that the actual water quantity is close to 90% of the target water quantity, and the prior parameter φ t of the current meta-network model is 0.5, then the updated prior parameter φ t+1 of the meta-network model is calculated as follows:

[0117]

[0118] The updated prior parameter φ t+1 of the meta-network model is calculated to be approximately equal to 0.5002421, the prior parameter of the meta-network model reflects the initial assumption or learning state of the meta-network model for different tasks, and the prior parameter φ t+1 of the meta-network model is updated with each feedback, and the prediction accuracy of the meta-network model is gradually improved.

[0119] The above merely provides a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0120] Finally: the above only for the preferred embodiments of the present application, and not for limiting the present application, any modification made within the spirit and principles of the present application, equivalent replacement, improvement, etc., should be included in the scope of protection of the present application.

Claims

1. An artificial intelligence-based irrigation water quantity prediction and control system, characterized by, Comprise: M1: heterogeneous data acquisition module, deploy sensors in irrigation area, and synchronously collect multi-dimensional environmental data through heterogeneous communication protocol; M2: collaborative repair module, construct sensor physical topology graph based on the multi-dimensional environmental data, detect hardware drift and signal conflict using spatio-temporal constraint graph attention network, and output spatio-temporal calibration data combined with causal reasoning loss function; M3: incremental prediction module, receive the spatio-temporal calibration data, generate scene adaptive parameters based on meta-network, fuse historical data distribution and real-time features through incremental learning, and output water quantity prediction results and confidence parameters using Bayesian neural network; M4: reinforcement control module, generate control instructions based on the water quantity prediction results and confidence parameters, and update meta-network prior parameters based on the feedback signals of instruction execution in reverse gradient.

2. The artificial intelligence-based irrigation water quantity prediction and control system according to claim 1, characterized in that, The sensors include rainfall sensors, temperature and humidity sensors, soil moisture sensors, water level sensors and flow sensors; Wherein, the rainfall sensor data symbol is represented as R(u,j); the temperature and humidity sensor data symbol is represented as T(u,j); the soil moisture sensor data symbol is represented as S(u,j); the water level sensor data symbol is represented as W(u,j); the flow sensor data symbol is represented as Q(u,j); wherein, u represents the time round of data collection; j represents the data collected by the corresponding sensor.

3. The artificial intelligence-based irrigation water quantity prediction and control system according to claim 2, characterized in that, The heterogeneous communication protocol includes LoRa protocol, NB-IoT protocol and 4G protocol, the LoRa protocol is suitable for the rainfall sensor and soil moisture sensor, the communication time slot length is 10 seconds; the NB-IoT protocol is suitable for the temperature and humidity sensor and flow sensor, the communication time slot length is 8 seconds; the 4G protocol is suitable for the water level sensor, the communication time slot length is 6 seconds.

4. The artificial intelligence-based irrigation water quantity prediction and control system according to claim 1, characterized in that, A physical topology of the sensors is constructed by spatial positions of the sensors to reflect relationships between the sensors, define geographical coordinates of the sensor p and the sensor q as (x p ,y p ) and (x q ,y q ) respectively, and a distance between the sensor p and the sensor q is D pq ; In the spatiotemporal constraint graph attention network, the relationship between the sensors is represented by a graph structure, and the data transmission between the sensors is weighted by using an attention mechanism, and the data collected by each sensor p at time t is x p,t , which represents the data characteristics of sensor p at time t, the attention weight α pq in the graph attention network reflects the degree of influence of sensor p on sensor q, and the calculation formula is: where exp(*) is the exponential function; LR(*) is the LeakyReLU activation function; b is the weight vector, representing the learning parameters related to the data of each sensor; b T is the transpose of the weight vector b, representing the conversion of its column vector to a row vector; W is the learning weight matrix, used to map the sensor data x p and x q to a new feature space; Wx p represents the linear transformation result of the features of sensor p; x p and x q represent the feature vectors of sensor p and sensor q at time t, respectively; ‖ is the concatenation operation; is the neighbor set of sensor p, representing the set of sensors directly connected to sensor p in the graph; Wx r is the linear transformation result of the features of neighbor sensor r; the linear transformation result Wx r by a weight matrix W and input data x of a sensor r r is calculated, and the formula is Wx r = W · x r wherein the weight matrix W represents a linear mapping of feature data of a sensor p in a sensor r; The attention weight a is calculated as pq For weighted fusion of the data of the sensor p, the formula is expressed as: wherein, is the data representation of the sensor p after the weighted fusion; data based on weighted fusion sensor prediction values of a graph attention network wherein W1 is a first layer weight matrix to perform linear transformation on input data ; W2 is a second layer weight matrix to perform processing on data processed by the activation function; b1 is a first layer bias term for data from the input layer to the hidden layer; b2 is a second layer bias term for data from the hidden layer to the final output layer; ReLU(*) is a ReLU activation function, representing ReLU(x) = max(0, x), to introduce a non-linear transformation, and x represents input data passed to the ReLU activation function.

5. The artificial intelligence-based irrigation water quantity prediction and control system according to claim 4, characterized in that, sensor actual measurement y t The actual observation data provided by the sensor, obtained directly through the sensor, the sensor prediction value The sensor actual measurement y t The error loss is calculated as follows: where L e is the error loss function of hardware drift and signal conflict, which is used to measure the difference between the actual measurement value of the sensor and the predicted value of the sensor; N is the length of the time series, representing the total length of the observation data; The causal reasoning loss function calibrates the sensor data by processing the causal relationship of spatio-temporal data using the minimum loss function, the formula is: wherein L ca is a loss function of the causal inference, representing the difference between the predicted value and the actual observed value; θ is a parameter of the causal inference model, representing the learning result of the model; f(x t , θ) is a predicted value generated by the causal inference model based on the input data x t and the parameter θ; is an L2 norm, representing the squared error between the predicted value and the actual value; Finally, the spatio-temporally calibrated sensor outputs the spatio-temporally calibrated data X' and X' = X1', X2',..., X p '.

6. The artificial intelligence-based irrigation water quantity prediction and control system according to claim 1, characterized in that, The meta-network generates adaptive parameters suitable for different tasks by learning from multiple tasks: wherein, is an optimal parameter, representing a parameter that causes a loss function to be minimized; represents finding a parameter that minimizes a loss function T is a task, representing a specific data processing task; E T～p(T) represents taking an expectation over a task distribution p(T); is a loss function for a task T, used to measure the prediction error of the meta- network model under the task; is a meta-network model output for a given meta-network parameter , a prediction result calculated based on an input data X' The incremental learning updates the parameters of the meta-network model by combining historical data and real-time data, the incremental learning formula is: wherein, denotes the meta-network model parameter at time t+1; denotes the meta-network model parameter at time t; a is a learning rate, used to control the magnitude of parameter update; is the gradient; The updated meta-network model uses the Bayesian neural network for water quantity prediction to provide confidence for each prediction value, reflecting the uncertainty of the prediction result: wherein, for a given input data X and meta-network model parameters a water quantity prediction result y output by the Bayesian neural network; denotes different values of the Bayesian neural network model parameters; is a Bayesian posterior distribution; is an integral of the Bayesian posterior distribution with respect to the Bayesian neural network model parameters ; The confidence parameter s is based on the Bayesian posterior distribution and the integral is calculated as Finally, output the water quantity prediction result y and the confidence parameter σ.

7. The artificial intelligence-based irrigation water quantity prediction and control system according to claim 6, characterized in that, generating the control instruction u based on the water amount prediction result y and the confidence parameter σ i The formula is expressed as follows: u i = c1 · y + c2 · σ + k Wherein, c1 and c2 are weight parameters, which are used to control the influence of water quantity prediction result and confidence parameter on control instruction respectively; k is a bias term, which is used to adjust the reference value of instruction; According to the control instruction u i The execution feedback defines the feedback signal as r i The gradient reverse update formula combined with the loss function is: wherein φ t+1 is the updated meta-network model prior parameter; φ t is the prior parameter of the current meta-network model; z is a learning rate for controlling the update amplitude; is the control instruction u i of the gradient of the meta-network model prior parameter φ t .

Images