A method for predicting and warning the frost damage risk of mountain orchard
Patent Information
- Application Number
- CN202310269864.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-20
- Publication Date
- 2026-01-09
- Estimated Expiration
- 2043-03-20
Smart Images
Figure CN116384737B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of frost damage risk monitoring, and more specifically, to a method for predicting and warning of frost damage risks in mountain orchards. Background Technology
[0002] Environmental monitoring is extremely important for the management of mountain orchards. This includes long-term monitoring of various environmental indicators related to the growth of fruit trees in mountain orchards, such as temperature, humidity, light, wind, pests and diseases, and soil nutrients.
[0003] In addition to monitoring the basic environment mentioned above, it is also necessary to monitor frost damage to fruit trees in mountain orchards. Once frost damage occurs, it will seriously affect the yield of the fruit.
[0004] Common causes of frost damage to fruit trees include: 1. Sudden temperature drop: During the transitional periods from late autumn to early winter and from late winter to early spring, the weather is unpredictable, with drastic and irregular temperature changes. For example, a sudden drop in temperature brought by a cold front in early winter, or a cold snap in early spring, can easily lead to frost damage to branches and trunks or the death of flower buds. 2. Sustained low temperatures: During the winter dormancy period, prolonged periods of low temperatures or excessively high temperatures can easily cause frost damage to fruit trees if growers fail to take preventative measures. 3. Large diurnal temperature range: In autumn, a large temperature range is beneficial for fruit tree expansion, sweetening, and color development. However, in winter, a large diurnal temperature range can easily lead to frost damage to the branches and trunks.
[0005] Therefore, effectively predicting and warning about frost damage risks in mountain orchards, so as to take frost protection measures for fruit trees in advance, is an important means to ensure fruit yield and avoid serious yield reduction. Summary of the Invention
[0006] The purpose of this application is to provide a method for predicting and warning of frost damage risk in mountain orchards, so as to provide early warning of frost damage risk in mountain orchards.
[0007] To achieve the above objectives, the embodiments of this application are implemented in the following manner:
[0008] In a first aspect, embodiments of this application provide a method for predicting and warning of frost damage risk in mountain orchards, including: obtaining data on the frost damage risk of mountain orchard areas in the past... Historical weather data for the day and in the future The forecast weather data for the day; based on the historical weather data and the forecast weather data, the input data is determined, wherein the input data includes... Historical weather feature vectors and a weather forecast feature vector; inputting the input data into a preset frost damage risk prediction model to obtain a frost damage risk prediction result output by the frost damage risk prediction model, and generating a frost damage risk prompt information based on the frost damage risk prediction result.
[0009] With reference to the first aspect, in a first possible implementation manner of the first aspect, based on the historical weather data and the weather forecast data, the input data is determined by: performing feature extraction on the historical weather data to obtain a historical weather feature vector, wherein each historical weather feature vector contains daily maximum temperature, daily minimum temperature, daily high temperature duration, daily low temperature duration, daily humidity information, daily light information, daily rainfall information, and daily wind speed information; performing feature extraction on the weather forecast data to obtain a weather forecast feature vector, wherein each weather forecast feature vector contains predicted maximum temperature, predicted minimum temperature, predicted high temperature duration, predicted low temperature duration, predicted humidity information, predicted light information, predicted rainfall information, and predicted wind speed information; and determining the input data based on the historical weather feature vector and the weather forecast feature vector.
[0010] With reference to the first possible implementation manner of the first aspect, in a second possible implementation manner of the first aspect, the frost damage risk prediction model includes an input layer, a prediction layer, and an output layer. The input layer is configured to receive the input data and input the input data to the prediction layer according to a set feature vector sequence input mode. The prediction layer includes a forward Lstm chain and a reverse Lstm chain. The forward Lstm chain includes a forward Lstm unit, configured to process the input feature vector sequence in a forward direction. The reverse Lstm chain includes a reverse Lstm unit, configured to process the input feature vector sequence in a reverse direction, so that the prediction layer determines a prediction vector. The output layer is configured to determine a frost damage risk prediction result based on the prediction vector and output the frost damage risk prediction result.
[0011] With reference to the second possible implementation manner of the first aspect, in a third possible implementation manner of the first aspect, the forward Lstm units of the forward Lstm chain are , respectively, and the reverse Lstm units of the reverse Lstm chain are , , When the first feature vector in the input feature vector sequence is , as input, and outputs the hidden layer output to ; when the i-th feature vector in the input sequence of feature vectors and the hidden layer output of as input, and outputs the hidden layer output to ; when the i-th feature vector in the input sequence of feature vectors and the hidden layer output of as input; when the i-th feature vector in the input sequence of feature vectors as input, and outputs the cell output to ; when the i-th feature vector in the input sequence of feature vectors and the cell output of as input, and outputs the cell output to ; when the i-th feature vector in the input sequence of feature vectors and the cell output of and the cell output of , determines the prediction vector as the final output.
[0012] With reference to the third possible implementation of the first aspect, in a fourth possible implementation of the first aspect, for the forward Lstm unit of the forward Lstm chain ,
[0013] when ,
[0014] ,
[0015] At that time, the input gate formula is:
[0016] ,
[0017] in, for The input gate, , and All The input gate weights, for The hidden layer output, The first feature vector in the input feature vector sequence 1 eigenvector for Input gate bias;
[0018] At that time, the forgetting gate formula is:
[0019] ,
[0020] At that time, the forgetting gate formula is:
[0021] ,
[0022] in, for The Gate of Oblivion , and All The weight of the forget gate for The hidden layer output, The first feature vector in the input feature vector sequence 1 eigenvector for Forget gate bias;
[0023] When the output gate formula is:
[0024] ,
[0025] When the output gate formula is:
[0026] ,
[0027] in, for The output gate, , and All The output gate weights, for The hidden layer output, The first feature vector in the input feature vector sequence 1 eigenvector for The output gate bias.
[0028] In conjunction with the fourth possible implementation of the first aspect, in the fifth possible implementation of the first aspect, Unit output satisfy:
[0029] ,
[0030] ,
[0031] ,
[0032] ,
[0033] in, for The unit output, ; for The overall weight, for The hidden layer output, Represents element-level product. For activation function, It is the hyperbolic tangent function. As an intermediate quantity, and for Update the candidate vector values in the data. and For candidate vector weights, This is the bias for the candidate vector.
[0034] In conjunction with the third possible implementation of the first aspect, in the sixth possible implementation of the first aspect, for the reverse Lstm unit of the reverse Lstm chain... ,
[0035] At that time, the input gate formula is:
[0036] ,
[0037] At that time, the input gate formula is:
[0038] ,
[0039] in, for input gate, , , and are input gate weights, is a cell output, is a cell output, and are the i-th and j-th feature vectors in the input feature vector sequence, respectively, is an input gate bias;
[0040] the forget gate formula is:
[0041] ,
[0042] the forget gate formula is:
[0043] ,
[0044] wherein, is a forget gate, , , and are forget gate weights, is a cell output, is a cell output, and are the i-th and j-th feature vectors in the input feature vector sequence, respectively, is a forget gate bias;
[0045] the output gate formula is:
[0046] ,
[0047] the output gate formula is:
[0048] ,
[0049] wherein, is The output gate, , , and All The output gate weights, for The unit output, for The unit output, and These are the th elements in the input feature vector sequence. The eigenvector and the eigenvector 1 eigenvector for Output gate bias.
[0050] In conjunction with the sixth possible implementation of the first aspect, in the seventh possible implementation of the first aspect, the inverse Lstm unit... Unit output satisfy:
[0051] ,
[0052] ,
[0053] ,
[0054] ,
[0055] in, for The unit output, ; for The overall weight, for The hidden layer output, Represents element-level product. For activation function, It is the hyperbolic tangent function. As an intermediate quantity, and for Update the candidate vector values in the data. and For candidate vector weights, This is the bias for the candidate vector.
[0056] Combining the third possible implementation of the first aspect, in the eighth possible implementation of the first aspect, based on Unit output and Unit output Determine the prediction vector for the final output. ,include:
[0057] The prediction vector is calculated using the following formula. :
[0058] ,
[0059] in, and All are weights. ;
[0060] The loss function is designed as follows:
[0061] ,
[0062] in, For loss function, For the sample size, For the first The predicted vector for each sample. For the first The true vector of each sample.
[0063] In conjunction with the second possible implementation of the first aspect, in the ninth possible implementation of the first aspect, the output layer determines the frost damage risk prediction result based on the prediction vector in the following way: the output layer calculates the frost damage risk score based on the prediction vector and the preset risk coefficient; and generates the frost damage risk prediction result based on the frost damage risk score.
[0064] Beneficial effects: This plan utilizes historical weather data (including...) Historical weather feature vectors) and forecast weather data (including Using forecast weather feature vectors as input data, a pre-defined frost damage risk prediction model is employed to obtain frost damage risk prediction results, enabling frost damage early warning for mountain orchards. Each weather feature vector encompasses factors affecting fruit tree frost damage, such as maximum temperature, minimum temperature, duration of high temperatures, duration of low temperatures, humidity, sunlight, rainfall, and wind speed. A targeted frost damage risk prediction model is designed to overcome the shortcomings of traditional prediction models that use historical data but lack forecast data. This allows for more accurate identification of frost damage risks, timely warnings, and prevention of frost damage to fruit trees. The use of a partially bidirectional Last-Stage Model (LSTM) structure (bidirectional LSTM for forecast weather data and unidirectional LSTM for historical data, allowing for consideration of historical factors while also emphasizing forecast weather data) significantly improves the reliability and accuracy of the prediction vectors.
[0065] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0066] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0067] Figure 1 This is a schematic diagram of the prediction layer in the frost damage risk prediction model provided in the embodiments of this application.
[0068] Figure 2 A flowchart illustrating a method for predicting and alerting the risk of frost damage in a mountain orchard, as provided in this application embodiment. Detailed Implementation
[0069] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.
[0070] To facilitate understanding of this scheme, the frost damage risk prediction model will be introduced first.
[0071] In this embodiment, a frost damage risk prediction model was designed to monitor the frost damage risk in mountain orchards. The frost damage risk prediction model may include an input layer, a prediction layer, and an output layer.
[0072] The input layer is used to receive input data in a specified format so that the feature vectors of the input data can be fed into the prediction layer according to the specified requirements.
[0073] For example, considering the data used in this scheme: the mountain orchard area in the past Historical weather data for the day and in the future The forecast weather data for the day. Therefore, when designing the frost damage risk prediction model, it is necessary to design a system that requires the use of... The prediction layer consists of feature vectors, and based on this, the input data is designed to satisfy the following format: Historical weather feature vectors and A total of 1 weather forecast feature vectors, The form of feature vectors (which can be arranged in chronological order) The feature vector set (each feature vector can contain multiple dimensions of attributes). After receiving the input data, the input layer can input the input data into the prediction layer according to the set feature vector sequence input pattern.
[0074] For example, the input data contains Historical weather feature vectors and A weather forecast feature vector, denoted as , The input layer can then... The input is fed into the corresponding part of the prediction layer, thereby realizing the input of the feature vector sequence.
[0075] Please see Figure 1 , Figure 1 This is a schematic diagram of the prediction layer in the frost damage risk prediction model provided in the embodiments of this application.
[0076] In this embodiment, the prediction layer may include a forward Lstm chain and a reverse Lstm chain. The forward Lstm chain includes... A forward Lstm unit is used to process the input feature vector sequence in the forward direction, while the reverse Lstm chain contains... Each inverse Lstm unit is used to process the input feature vector sequence in reverse. Based on the processing of the feature vector sequence by the forward and inverse Lstm chains, the prediction layer can determine the prediction vector.
[0077] For example, a forward Lstm chain The positive Lstm units are respectively , .
[0078] hour, (Right now The first feature vector in the input feature vector sequence As input, and output the hidden layer (Right now Output to .
[0079] hour, The first feature vector in the input feature vector sequence eigenvectors and Hidden layer output As input, and output the hidden layer Output to .
[0080] hour, (Right now ) The first feature vector in the input feature vector sequence eigenvectors and Hidden layer output As input.
[0081] Specifically, for the forward Lstm unit of the forward Lstm chain :
[0082] At that time, the input gate formula is:
[0083] , (1)
[0084] At that time, the input gate formula is:
[0085] , (2)
[0086] in, for The input gate, , here The independent variable is (in Formula 1, the independent variable is...) In Formula 2, the independent variable is ), and All The input gate weights, for The hidden layer output, The first feature vector in the input feature vector sequence 1 eigenvector for Input gate bias.
[0087] At that time, the forgetting gate formula is:
[0088] , (3)
[0089] At that time, the forgetting gate formula is:
[0090] , (4)
[0091] in, for The Gate of Oblivion , and All The weight of the forget gate for The hidden layer output, The first feature vector in the input feature vector sequence 1 eigenvector for The forget gate bias.
[0092] When the output gate formula is:
[0093] , (5)
[0094] When the output gate formula is:
[0095] , (6)
[0096] in, for The output gate, , and All The output gate weights, for The hidden layer output, The first feature vector in the input feature vector sequence 1 eigenvector for Output gate bias.
[0097] as well as, Unit output satisfy:
[0098] , (7)
[0099] , (8)
[0100] (9)
[0101] , (10)
[0102] in, for The unit output, ; for The overall weight, for The hidden layer output, Represents element-level product. For activation function, It is the hyperbolic tangent function. As an intermediate quantity, and for Update the candidate vector values in the data. and For candidate vector weights, This is the bias for the candidate vector.
[0103] An example, a reverse Lstm chain The inverse Lstm units are respectively , .
[0104] When, the first feature vector in the input feature vector sequence is inputted, the cell output is outputted.
[0105] When, the first feature vector and the second feature vector in the input feature vector sequence are inputted, the cell output is outputted.
[0106] When, the first feature vector and the second feature vector in the input feature vector sequence
[0107] are inputted. The reverse LSTM unit of the reverse LSTM chain:
[0108] When, the input gate formula is:
[0109] , (11)
[0110] When, the input gate formula is:
[0111] , (12)
[0112] wherein, is the input gate of , , , and are the input gate weights of , is the cell output of , is the cell output of , and are the first feature vector and the second feature vector in the input feature vector sequence, is the input gate bias of .
[0113] At that time, the forgetting gate formula is:
[0114] , (13)
[0115] At that time, the forgetting gate formula is:
[0116] (14)
[0117] in, for The Gate of Oblivion , , and All The weight of the forget gate for The unit output, for The unit output, and These are the th elements in the input feature vector sequence. The eigenvector and the eigenvector 1 eigenvector for The forget gate bias.
[0118] When the output gate formula is:
[0119] , (15)
[0120] When the output gate formula is:
[0121] (16)
[0122] in, for The output gate, , , and All The output gate weights, for The unit output, for The unit output, and These are the th elements in the input feature vector sequence. The eigenvector and the eigenvector 1 eigenvector for Output gate bias of the LSTM unit.
[0123] Reverse LSTM unit Unit output of the LSTM unit Satisfies:
[0124] , (17)
[0125] , (18)
[0126] , (19)
[0127] , (20)
[0128] Wherein, is Unit output of the LSTM unit, ; is Integrated weight of the LSTM unit, is Hidden layer output of the LSTM unit, represents element-level product, is an activation function, is a hyperbolic tangent function, is an intermediate quantity, and is Candidate vector update value in the LSTM unit, and are candidate vector weights, is a candidate vector bias.
[0129] For the weights in the scheme (for example and , and , and , , and , , and , , and , , and , , and , etc.), the weight is composed of S vectors The weight vector can be regularized by using the maximum norm constraint. The constraint coefficient ,when season Then let each Multiply Normalize its norm to a constant. .
[0130] based on Unit output and Unit output This allows us to determine the prediction vector for the final output. .
[0131] Specifically, the prediction vector can be calculated using the following formula. :
[0132] , (twenty one)
[0133] in, and All are weights. .
[0134] The loss function is designed as follows:
[0135] , (twenty two)
[0136] in, For loss function, For the sample size, For the first The predicted vector for each sample. For the first The true vector of each sample.
[0137] The output layer of the frost damage risk prediction model can determine and output the frost damage risk prediction results based on the prediction vector.
[0138] For example, the output layer can calculate a frost damage risk score based on the prediction vector and a preset risk coefficient. Here, the prediction vector is a... dimensional vectors (e.g.) dimensional row vectors or If the predicted vector is a column vector of dimension 1, then the method for calculating the frost damage risk score can be: combine the predicted vector with a set vector (which is a column vector of dimension 2). dimensional column vector or The inner product of the dimensional row vectors (where the elements in the vector are determined based on the degree of influence of different elements in the prediction vector on the risk of frost damage) is calculated to obtain the frost damage risk score.
[0139] Then, the freeze risk score can be compared with a set freeze risk threshold (or according to the interval in which the freeze risk score falls), to determine a freeze risk prediction result, and output.
[0140] For the specific design of the freeze risk prediction model, the above has been introduced clearly. After the freeze risk prediction model is constructed, relevant data (for example, containing the historical weather data and the forecast weather data of the orchard location in the past 10 years, the weather data containing the highest temperature, the lowest temperature, the high temperature duration, the low temperature duration, the humidity information, the illumination information, the rainfall information and the wind speed information, etc. Information every day) can be obtained from the national comprehensive meteorological information sharing platform, and the data set required for training the model (for example, 3000 data are obtained, m is 10, n is 5, 2986 samples can be constructed, then the samples are divided into training set, validation set and test set according to 7:2:1, or the weather data from November to April of the next year is obtained, and the training set, the validation set and the test set are constructed on the basis of this), the constructed freeze risk prediction model is trained, until the training requirement is met, the training of the freeze risk prediction model is completed, and the trained freeze risk prediction model is obtained.
[0141] After that, the freeze risk prediction and alarm method for mountain orchard can be run based on the trained freeze risk prediction model.
[0142] Please refer to Figure 2 , Figure 2 The flow chart of a freeze risk prediction and alarm method for mountain orchard provided in the embodiment of the application.
[0143] In the embodiment, the running program of the freeze risk prediction and alarm method for mountain orchard is carried in the server, and the method can include steps S10, S20 and S30.
[0144] In order to predict the freeze risk of the next period (for example, the next day), the server can execute step S10.
[0145] Step S10: obtaining the historical weather data of the mountain orchard area in the past days and the forecast weather data in the future days.
[0146] In the embodiment, the server can obtain the historical weather data of the mountain orchard area in the past days and the forecast weather data in the future days from the national comprehensive meteorological information sharing platform (or other more accurate weather data sources), each weather data containing the highest temperature, the lowest temperature, the high temperature duration, the low temperature duration, the humidity information, the illumination information, the rainfall information and the wind speed information, etc. Information every day.
[0147] obtain historical weather data of the mountain orchard area in the past obtain forecast weather data of the mountain orchard area in the future After obtaining the historical weather data and the forecast weather data, the server can perform step S20.
[0148] Step S20: Based on the historical weather data and the forecast weather data, determine input data, wherein the input data comprises a historical weather feature vector, and a forecast weather feature vector.
[0149] In this embodiment, the server can perform feature extraction on the historical weather data to obtain a historical weather feature vector, wherein each historical weather feature vector comprises daily maximum temperature, daily minimum temperature, daily high temperature duration, daily low temperature duration, daily humidity information, daily light information, daily rainfall information, and daily wind speed information. Here, each historical weather feature vector is taken as an example of a row vector.
[0150] In addition, the server can perform feature extraction on the forecast weather data to obtain a forecast weather feature vector, wherein each forecast weather feature vector comprises predicted maximum temperature, predicted minimum temperature, predicted high temperature duration, predicted low temperature duration, predicted humidity information, predicted light information, predicted rainfall information, and predicted wind speed information. Here, each forecast weather feature vector is also taken as an example of a row vector.
[0151] Then, the server can determine input data based on the historical weather feature vector, and the forecast weather feature vector. For example, the historical weather feature vector and the forecast weather feature vector are constructed into a vector group (denoted as , …, , …, ) in chronological order as input data.
[0152] After that, the server can perform step S30.
[0153] Step S30: Input the input data into a preset frost damage risk prediction model to obtain a frost damage risk prediction result output by the frost damage risk prediction model, and generate a frost damage risk prompt information based on the frost damage risk prediction result.
[0154] In the embodiment, the server can input the input data into the trained freeze risk prediction model, process the input data (a sequence of feature vectors) by the freeze risk prediction model, the processing process can refer to the introduction of the freeze risk prediction model, and the details are not repeated here, and finally output a freeze risk prediction result. The server can generate freeze risk prompt information based on the freeze risk prediction result to prompt the user to take freeze protection in time.
[0155] In summary, the application provides a freeze risk prediction and warning method for mountain orchards, which uses historical weather data (containing historical weather feature vectors) and forecast weather data (containing forecast weather feature vectors) as input data, uses a preset freeze risk prediction model to obtain a freeze risk prediction result, and realizes freeze warning for mountain orchards. Each weather feature vector covers the factors affecting the occurrence of freeze damage of fruit trees, such as maximum temperature, minimum temperature, high temperature duration, low temperature duration, humidity information, light information, rainfall information, and wind speed information. The freeze risk prediction model is designed specifically to overcome the drawbacks of traditional prediction models that use historical data but lack forecast data, and can more accurately identify freeze risk and give timely reminders to prevent freeze damage of fruit trees. The use of a part of the bidirectional Lstm structure (bidirectional Lstm for forecast weather data and unidirectional Lstm for historical data) can greatly improve the reliability and accuracy of the prediction vector while giving priority to the forecast weather data. Moreover, the designed bidirectional Lstm structure can combine forward and reverse prediction of the forecast weather data to further improve the prediction accuracy.
[0156] In this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions.
[0157] The above only describes the embodiments of the application and does not limit the protection scope of the application. For those skilled in the art, the application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.
Claims
1. A method for predicting and warning the frost risk of a mountain orchard, characterized in that, The method comprises the following steps: acquire historical weather data for the mountain orchard area for the past days and forecasted weather data for the future days; Based on the historical weather data and the forecast weather data, input data is determined, wherein the input data comprises a historical weather feature vector and a forecast weather feature vector. inputting the input data into a preset frost damage risk prediction model to obtain a frost damage risk prediction result output by the frost damage risk prediction model, and generating a frost damage risk prompt information based on the frost damage risk prediction result; determining input data based on the historical weather data and the forecast weather data, comprising: feature extraction is performed on the historical weather data to obtain a plurality of historical weather feature vectors, wherein each historical weather feature vector contains a daily maximum temperature, a daily minimum temperature, a daily high temperature duration, a daily low temperature duration, daily humidity information, daily light information, daily rainfall information, and daily wind speed information; characteristic extraction is performed on the weather forecast data to obtain a weather forecast characteristic vector, wherein each weather forecast characteristic vector contains predicted maximum temperature, predicted minimum temperature, predicted high temperature duration, predicted low temperature duration, predicted humidity information, predicted light information, predicted rainfall information, and predicted wind speed information. based on a historical weather feature vector and a forecast weather feature vector, the input data is determined; The frost damage risk prediction model comprises an input layer, a prediction layer and an output layer, The input layer is used to receive the input data and input the input data into the prediction layer according to a set feature vector sequence input mode; The prediction layer includes a forward LSTM chain and a backward LSTM chain, the forward LSTM chain includes a plurality of forward LSTM units for forward processing the input sequence of feature vectors, and the backward LSTM chain includes a plurality of backward LSTM units for backward processing the input sequence of feature vectors, so that the prediction layer determines a prediction vector. The output layer is used to determine a frost damage risk prediction result based on the prediction vector and output the frost damage risk prediction result; The forward Lstm chain has forward Lstm units respectively , The reverse Lstm chain has reverse Lstm units respectively , Time, The first feature vector in the input sequence of feature vectors as input and the hidden layer output is output to ; hour, The first feature vector in the input feature vector sequence eigenvectors and Hidden layer output As input, and output the hidden layer Output to ; At time, the hidden layer output of the th feature vector in the input sequence of feature vectors and the hidden layer output are taken as input; time, with the input sequence of feature vectors the input sequence of feature vectors output of the cell ; hour, The first feature vector in the input feature vector sequence eigenvectors and Unit output Output to As input, and output the unit. Output to ; hour, The first feature vector in the input feature vector sequence eigenvectors and Unit output As input; Based on the unit outputs and the unit outputs , a prediction vector for the final output is determined .
2. The method for frost risk prediction and warning of mountain orchard according to claim 1, characterized in that, forward lstm unit of a forward lstm chain , The input gate formula is: ， The input gate formula is: ， wherein, is an input gate, , and are input gate weights of is a hidden layer output of is the th feature vector in the sequence of input feature vectors, is an input gate bias of . The forget gate formula is: ， The forget gate formula is: ， wherein, is the forget gate, , and are forget gate weights, is the hidden layer output, is the th feature vector in the input sequence of feature vectors, is the forget gate bias; The output gate formula is: ， The output gate formula is: ， in, for The output gate, , and All The output gate weights, for The hidden layer output, The first feature vector in the input feature vector sequence 1 eigenvector for The output gate bias.
3. The method for frost risk prediction and warning of mountain orchard according to claim 2, characterized in that, the unit output of the cell satisfies: ， wherein is a unit output, ; is a combined weight, is a hidden layer output, denotes an element-wise multiplication, is an activation function, is a hyperbolic tangent function, is an intermediate quantity, and is a candidate vector update value, and are candidate vector weights, is a candidate vector bias.
4. The frost risk prediction and warning method for mountain orchard according to claim 1, characterized in that, Reverse lstm unit for reverse lstm chain , The input gate formula is: ， The input gate formula is: ， in, for The input gate, , and All The input gate weights, for The unit output, for The unit output, and These are the first and second digits of the input feature vector sequence, respectively. The eigenvector and the eigenvector 1 eigenvector for Input gate bias; The forget gate formula is: ， The forget gate formula is: ， in, for The Gate of Oblivion , and All The weight of the forget gate for The unit output, for The unit output, and These are the first and second digits of the input feature vector sequence, respectively. The eigenvector and the eigenvector 1 eigenvector for Forget gate bias; The output gate formula is: ， The output gate formula is: ， in, for The output gate, , and All The output gate weights, for The unit output, for The unit output, and These are the th elements in the input feature vector sequence. The eigenvector and the eigenvector 1 eigenvector for The output gate bias.
5. The method for frost risk prediction and warning of mountain orchard according to claim 4, characterized in that, A reverse lstm unit The cell output of the unit Satisfies: ， wherein is the unit output, ; ; is the combined weight, ; is the hidden layer output, ; denotes an element-wise multiplication, is an activation function, is a hyperbolic tangent function, is an intermediate quantity, and is the candidate vector update value, ; and are the candidate vector weights, is the candidate vector bias.
6. The frost risk prediction and warning method for mountain orchard according to claim 1, characterized in that, based on the unit output and the unit output , determine the final output prediction vector , comprising: The prediction vector is calculated using the following equation : ， wherein and are weights, ; The loss function is designed as: ， wherein, is a loss function, is a number of samples, is a predicted vector for the th sample, is a true vector for the th sample.
7. The frost risk prediction and warning method for mountain orchard according to claim 1, characterized in that, The output layer determines the frost damage risk prediction result based on the prediction vector in the following manner: The output layer calculates a frost damage risk score based on the prediction vector and a preset risk coefficient; generate a frost damage risk prediction result based on the frost damage risk score.
Citation Information
Patent Citations
CNN-LSTM short-term wind power prediction method based on attention mechanism
CN113642225A