A Method and System for Predicting Moisture Content of Combustible Materials Based on Long Short-Term Memory Model

Abstract

This invention discloses a method and system for predicting the moisture content of combustibles based on a Long Short-Term Memory (LSTM) model. First, data is collected and processed, combining data sources for both combustible moisture content and meteorological data, and processing the data according to the measurement time and location. An LSTM model is then constructed and trained, and finally, the trained LSTM model is used for result prediction. This method can effectively predict the moisture content of combustibles. By training and validating the dataset using the LSTM model, the accuracy is improved compared to traditional machine learning algorithms. Furthermore, combustible moisture content has a high correlation with meteorological factors; this method integrates meteorological factors such as temperature, vapor pressure, and wind speed, further improving the prediction accuracy compared to using only combustible moisture content for prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing, and particularly relates to a combustible moisture content prediction method and system based on a long short-term memory model. BACKGROUND

[0002] Wildfire plays an important role in regulating global carbon cycle of ecological system, changing vegetation succession mode, resisting disease and insect pests, and other ecological system natural evolution and development. However, wildfire is also a regional, even global natural disaster, which can destroy ecological system, release greenhouse gases, pollute the atmosphere, cause casualties and property losses. On November 8, 2018, a wildfire broke out in the town of Paradise, Butte County, northern California, USA. This wildfire has become the most destructive fire in the history of California. The burned area exceeds 150,000 acres, and the nearest Paradise town with only 27,000 population was almost burned into a ruin. Therefore, it is very important to develop a wildfire risk assessment and early warning system for maintaining ecological environment stability, protecting people's life and property safety, and achieving the sustainable development goals of the United Nations. For wildfire risk assessment, Pyne et al. proposed a famous wildfire triangle model in 1996. The model points out that meteorology, topography and fuel are three important indicators for assessing wildfire risk and spreading trend.

[0003] Fuel information includes fuel moisture content, fuel load and fuel category. Since vegetation leaves are more likely to be ignited than vegetation branches, and the total surface area of vegetation leaves is much larger than that of vegetation branches, in the study of wildfire, fuel refers to leaves, i.e. vegetation canopy. Among them, LFMC refers to vegetation canopy fuel moisture content, i.e. live fuel moisture content, which is defined as the ratio of vegetation canopy water content to its dry weight. How to obtain accurate FMC is the key to wildfire risk assessment and early warning.

[0004] Currently, there are mainly three ways to obtain fuel moisture content: field measurement, empirical model-based acquisition and physical model-based inversion method. The physical model method describes the explicit physical mechanism between vegetation fuel moisture content and canopy reflectivity through a physical model, and has higher universality, but has higher complexity and low model operation efficiency. Moreover, the physical model method can only rely on remote sensing data to obtain near real-time fuel moisture content, and to obtain future fuel moisture content, at present, an empirical model based on historical and future meteorological data is used, but the effect is poor.

[0005] The combustible moisture content prediction method based on the long short-term memory model is a prediction method with good effect and stability. At present, the main steps of the combustible moisture content prediction method based on the long short-term memory model are as follows: firstly, relevant combustible moisture content data and meteorological data are downloaded; secondly, functions in the long short-term memory model are determined and adjusted, including window length, activation function, error calculation function and iterative updating mode, forward calculation and back propagation training are carried out; finally, the data set is randomly divided into a training set and a validation set according to a certain proportion to prevent overfitting. Although this method has good effect, it still has many problems. First, the structure of the long short-term memory model is relatively complex and depends on the original data set, which cannot guarantee the adaptation to all regions and vegetation. Second, when selecting meteorological data and other auxiliary characteristic factors, the effect will change to different degrees with the selection of these factors. Therefore, there is still great limitation in predicting combustible moisture content in a large range. SUMMARY

[0006] Therefore, the purpose of the present application is to provide a combustible moisture content prediction method and system based on a long short-term memory model, which uses a long short-term memory model to predict combustible moisture content, and determines various functions in the neural network model to reasonably predict combustible moisture content by comprehensively using meteorological data such as air temperature, water vapor pressure and rainfall.

[0007] To achieve the above purpose, the present application provides the following technical scheme:

[0008] The combustible moisture content prediction method based on the long short-term memory model provided by the present application comprises the following steps:

[0009] Obtain combustible moisture content data and meteorological data, wherein the combustible moisture content data includes a fuel moisture database and measured values of combustible moisture content, and the measured values of combustible moisture content are used as basic characteristic variables; the meteorological data are used as auxiliary characteristic variables;

[0010] Construct a long short-term memory model, which is used to selectively store data at each time;

[0011] According to the combustible moisture content data and the meteorological data, a training set and a validation set of the long short-term memory model are extracted, the long short-term memory model is trained through the training set, and the validation set data are predicted through the trained long short-term memory model, and a prediction result is output.

[0012] Further, the meteorological data include maximum temperature, minimum temperature, average temperature, dew point temperature, maximum water vapor pressure, minimum water vapor pressure and wind speed, and data processing is performed according to the measurement time and measurement place of the combustible moisture content;

[0013] Further, the training process of the long short-term memory model is performed according to the following steps:

[0014] The window length of the training set is determined, the activation function of the long short-term memory module and the activation function of the output are determined, the discard rate of each layer of network nodes is determined, the error calculation method and the iterative updating method of the weight parameters are determined, and the cycle number and batch size of the model training are determined.

[0015] Further, the combustible water content and the meteorological data adopt the time sequence determined by the US National Fuel Moisture Database.

[0016] Further, the long short-term memory model comprises a recurrent neural network model and a cell state, the cell state is arranged in a hidden layer in the recurrent neural network model, the cell state is used to save the long-term state of the model, the long-term state is controlled through a gating device, and the gating device comprises a forget gate, an input gate and an output gate.

[0017] The forget gate is used to determine how much of the cell state at the last moment is retained to the current moment.

[0018] The input gate is used to determine how much of the input of the network at the current moment is saved to the cell state.

[0019] The output gate is used to control the output of the information of the cell state to the current output value of the long short-term memory model.

[0020] The combustible water content prediction system based on the long short-term memory model provided by the present application comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the computer program is executed by the processor to realize the above-mentioned method.

[0021] The present application has the following advantages:

[0022] The combustible moisture content prediction method and system based on the long short-term memory model provided by the application first collects and processes data, uses the national fuel moisture database of the United States as the data source of the combustible moisture content, and uses the measured value of the combustible moisture content as the basic characteristic variable; using PRISM climate group as the data source of meteorological data, including maximum temperature, minimum temperature, average temperature, dew point temperature, maximum water vapor pressure, minimum water vapor pressure and wind speed and other 7 auxiliary characteristic variables, and data processing according to the measurement time and measurement place of the combustible moisture content. Then determine and adjust the long short-term memory model, generate the model training data set, determine the window length of the training set; determine the activation function of the long short-term memory module and the activation function of the output; in order to prevent overfitting, determine the discard rate of each layer of network nodes; determine the error calculation method and the iterative updating method of the weight parameter; determine the cycle number and batch size of model training, and carry out forward calculation and back propagation training of the model. Finally, the long short-term memory model is trained and the result is predicted, the data set is randomly divided into training set and validation set according to a certain proportion, and overfitting is prevented. The method can effectively predict the combustible moisture content, train and verify the data set through the long short-term memory model, and the accuracy is improved compared with the traditional machine learning algorithm; and the combustible moisture content and meteorological factors have high correlation, and the method integrates temperature, water vapor pressure, wind speed and other meteorological factors, and the prediction accuracy is further improved compared with only using the combustible moisture content. Compared with the ordinary recurrent network model, the method can process long-distance dependence problem, so that the long-term and short-term inputs are comprehensively applied, and the time sequence characteristics of the combustible moisture content can be well predicted, which provides good data support for fire risk early warning.

[0023] Other advantages, objects, and features of the present application will be understood by those skilled in the art from the following specification in conjunction with the accompanying drawings. The present application's objects and other advantages will be realized and attained by the below specification. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to make the objects, technical solutions and beneficial effects of the present application clearer, the present application provides the following drawings for explanation:

[0025] Figure 1 The combustible moisture content prediction method flow chart based on the long short-term memory model.

[0026] Figure 2 The principle diagram of the recurrent neural network (RNN).

[0027] Figure 3 The long short-term memory model gate schematic diagram.

[0028] Figure 4 Loss diagram for training and verification of long short-term memory model.

[0029] Figure 5 Time series of predicted fuel moisture content (FMC) and measured fuel moisture content (FMC) for the model. DETAILED DESCRIPTION

[0030] The application will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the application and implement it. The embodiments are not intended to limit the application.

[0031] Embodiment 1

[0032] As shown in Figure 1 , Figure 1 Flow chart of fuel moisture content prediction method based on long short-term memory model, the fuel moisture content prediction method based on long short-term memory model provided in this embodiment, by constructing a long short-term memory model, taking fuel moisture content as a basic characteristic variable, and comprehensively considering auxiliary characteristic variables such as air temperature, dew point temperature, water vapor pressure and precipitation, a method for predicting fuel moisture content is provided, and the specific steps are as follows:

[0033] Obtain fuel moisture content data and meteorological data, the fuel moisture content data including a fuel moisture database and measured values of fuel moisture content, the measured values of fuel moisture content serving as basic characteristic variables; the meteorological data serving as auxiliary characteristic variables;

[0034] Construct a long short-term memory model, the long short-term memory model including an input layer, a hidden layer and an output layer, in the hidden layer, unit states and three kinds of gate devices are set to selectively store information at previous time, so that information at each time will contain information at the previous time, and the information at the previous time is selectively discarded. The three kinds of gate devices are respectively a forget gate, an input gate and an output gate. The long short-term memory model can be used for simulation and prediction of time series information.

[0035] According to the fuel moisture content data and the meteorological data, a training set and a verification set of the long short-term memory model are extracted, the long short-term memory model is trained through the training set, and the long short-term memory model after training is used to predict the verification set data, and the prediction result is output.

[0036] The combustible moisture content has obvious time sequence through a large number of measured data observation. As a special recurrent neural network, the long short-term memory model is very effective for data with sequence characteristics, which can mine the time sequence information and semantic information in the center of data. The ordinary recurrent neural network model will have the problems of gradient explosion and disappearance, and the long short-term memory model can effectively alleviate this problem through the "gate device" to selectively store information.

[0037] As Figure 2 shown, Figure 2 is the schematic diagram of recurrent neural network (RNN). The left figure is the recurrent neural network diagram, and the right figure is the RNN unfolded into a full neural network on the time coordinate axis. x t represents the input at time t; s t is the unit state of the hidden layer at time t, U is the connection matrix of the input layer, and W is the weight matrix of the hidden layer at the previous time to the hidden layer at the next time; o t is the output at time t, and V is the connection matrix of the output layer.

[0038] The calculation process of recurrent neural network is as follows:

[0039] o t = g (V * s t ) (1)

[0040] s t = f (U * x t + W * s t-1 ) (2)

[0041] Where x t represents the input at step t; s t is the state of the hidden layer at step t, which is the memory unit of the network, and it is calculated according to the input of the current input layer and the state of the hidden layer at the previous time. As shown in equation (2). Where U is the connection matrix of the input layer, W is the weight matrix of the hidden layer at the previous time to the hidden layer at the next time, f(x) is generally a nonlinear activation function, o t is the output at step t, and the output layer is a fully connected layer, that is, each node of the output layer is connected to each node of the hidden layer, and V is the connection matrix of the output layer, and g(x) is the activation function. As can be seen, the output value of the recurrent neural network is related to the historical input value at the previous multiple times.

[0042] As Figure 3 shown, Figure 3 is the schematic diagram of long short-term memory model gate. Including forget gate, input gate and output gate. h t-1 and x trepresents the output value of the LSTM at the previous time step and the input value at time step t, respectively. st is the cell state of the hidden layer at time step t, and σ is the sigmoid function.

[0043] Long Short-Term Memory (LSTM) models add an extra state in the hidden layers compared to recurrent neural network (RNN) models. This extra state is called the cell state and is used to store long-term states. The model controls the long-term state through three switches: the forget gate, the input gate, and the output gate.

[0044] The forget gate determines how much of the cell's state from the previous time step is retained in the current time step; w f It is the weight matrix of the forget gate, h t-1 and x t These represent the output value of the LSTM at the previous time step and the input value at the current time step, respectively. f σ is the bias term of the forget gate, and σ is the sigmoid function.

[0045] f t =σ(W f *[h t-1 ,x t ]+b f (3)

[0046] The input gate determines how much of the network's current input is stored in the cell state; i t This indicates the input to the cell at this moment.

[0047] In the state, This indicates that the tanh function is used to describe the current input cell state, s t f is obtained by multiplying the previous cell state element-wise by the forget gate. t Then use the current input cell state s t i obtained by multiplying the element by the input gate t The two are added together to obtain a new cell state.

[0048] i t =σ(W i *[h t-1 ,x t ]+b i (4)

[0049]

[0050]

[0051] Output gate, control unit status s t How much output is added to the current output value h of the Long Short-Term Memory model? t middle.

[0052] o t=σ(W o *[h t-1 ,x t ]+b o (7)

[0053] h t =o t *tanh(s t (8)

[0054] Example 2

[0055] The Long Short-Term Memory (LSTM) model provided in this embodiment comprises three parts: an input layer, a hidden layer, and an output layer. In the hidden layer, cell states and three gating mechanisms are used to selectively store information from previous time steps, ensuring that each time step contains information from the previous time step while selectively discarding some of that information. The three gating mechanisms are the forget gate, the input gate, and the output gate. The cell states are used to store the long-term state of the model, which is controlled by three switches: the forget gate, the input gate, and the output gate. The forget gate determines how much of the cell state from the previous time step is retained in the current time step; the input gate determines how much of the network's input at the current time step is stored in the cell states; and the output gate controls how much of the cell state is output to the current output value of the LTM model.

[0056] The specific steps of the method for predicting the moisture content of combustibles using a long short-term memory model are as follows:

[0057] (1) Data preparation

[0058] Fuel Moisture Content (FMC) refers to the ratio of the difference between the wet weight and the dry weight of plants in a sampling unit to the dry weight. It reflects the water content of a unit of vegetation leaf and its level directly affects the fire hazard level. It is a very important fire inducing factor and can be obtained through inversion using a vegetation radiative transfer model.

[0059] In this embodiment, the moisture content of combustible materials and various meteorological data are selected as training data for the long short-term memory model. The meteorological data includes maximum temperature, minimum temperature, average temperature, maximum vapor pressure, minimum vapor pressure, rainfall, and wind speed. These factors are used as feature variables to detail the implementation scheme, with California, USA, as the case study.

[0060] (2) Determine the model training dataset

[0061] like Figure 4 As shown, Figure 4 This diagram illustrates the loss during training and validation of a Long Short-Term Memory (LSTM) model. The diagram shows the number of training iterations and the final error result.

[0062] Figure 5 The figure shows the time series of model-predicted and measured combustible moisture content (FMC). The difference between the model-predicted FMC and the measured FMC on the time axis is compared in the figure.

[0063] Meteorological data is converted to the same format as the moisture content of combustibles using data processing methods. The window length of the training set is determined to be 3, meaning that the moisture content of combustibles can be predicted after 3 units.

[0064] (3) Forward computation

[0065] according to Figure 3 The output value of each neuron is calculated forward, involving a total of 5 variables. The calculation method is shown in the principle formulas (1)-(8). t This indicates how many cell states from the previous time step are retained in the current time step; i t This indicates how much of the network's input is currently stored in the cell state; Used to describe the current input cell state; s t Indicates the current cell state; h t This represents the current output value of the LSTM.

[0066] (4) Backpropagation training algorithm.

[0067] The error term value for each neuron is calculated in reverse. One method is backward propagation along time, that is, starting from the current time t, calculating the error term at each time step, and the other is propagating the error term up one level.

[0068] (5) Calculate the gradient of each weight based on the corresponding error term.

[0069] Based on the above calculations, the ultimate goal is to learn 8 sets of parameters, including the weight matrix and bias term of the forget gate, the weight matrix and bias term of the input gate, the weight matrix and bias term of the output gate, and the weight matrix and bias term of the computation unit state.

[0070] (6) Divide the data into training and validation sets, and determine the cyclic batch and batch size for model training.

[0071] By training the model using different cyclic batches, we can calculate the model error. The error function uses the MSE-loss function. Once the cyclic batch reaches a certain size, the error value will not change. We ultimately select the cyclic batch with the best performance.

[0072] (7) Use the trained model to conduct further experiments on the validation set.

[0073] The validation set is input into the trained model, and the model's predictions can be obtained through forward computation.

[0074] The combustible moisture content prediction system based on the long short-term memory model provided by the embodiment comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the computer program realizes the above method when executed by the processor.

[0075] The above-described embodiments are only preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Any equivalent replacement or transformation made by those skilled in the art based on the present application is within the protection scope of the present application. The protection scope of the present application is subject to the claims.

Claims

1. A combustible moisture content prediction method based on a long short-term memory model, characterized by: The method comprises the following steps: obtaining combustible moisture content data and meteorological data, the combustible moisture content data including a fuel moisture database and measured values of combustible moisture content as basic characteristic variables; the meteorological data as auxiliary characteristic variables; the meteorological data including maximum temperature, minimum temperature, average temperature, dew point temperature, maximum water vapor pressure, minimum water vapor pressure and wind speed, data processing according to the measurement time and measurement location of combustible moisture content; constructing a long short-term memory model for selectively storing data at each time; extracting a training set and a verification set of the long short-term memory model according to the combustible moisture content data and the meteorological data, training the long short-term memory model through the training set, and predicting the verification set data through the trained long short-term memory model to output a prediction result; the long short-term memory model including a recurrent neural network model and a cell state, the cell state being arranged in a hidden layer in the recurrent neural network model, the cell state being used to save a long-term state of the model, the long-term state being controlled through a gating device, the gating device including a forget gate, an input gate and an output gate; the forget gate being used to determine how much of the cell state at the previous time is retained to the current time; the input gate being used to determine how much of the input of the network at the current time is saved to the cell state; the output gate being used to control the output of the information of the cell state to the current output value of the long short-term memory model. 2.The combustible moisture content prediction method based on a long short-term memory model according to claim 1, wherein: the training process of the long short-term memory model being performed according to the following steps: determining the window length of the training set; determining the activation function of the long short-term memory module and the activation function of the output; determining the discard rate of each layer of network nodes; determining the error calculation method and the iterative update method of the weight parameters; determining the cycle number and batch size of the model training. 3.The combustible moisture content prediction method based on a long short-term memory model according to claim 1, wherein: the combustible moisture content data and the meteorological data adopting time series determined by the US National Fuel Moisture Database.

4. A combustible moisture content prediction system based on long short-term memory model, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that, the computer program being executed by the processor to realize the method according to any one of claims 1 to 3.

Images