A multi-source same type meteorological element fusion verification method

Abstract

The application discloses a multi-source same-type meteorological element fusion verification method, introduces target element characteristics and strongly related meteorological element characteristics of same-type stations, obtains a multi-feature fusion data set, introduces an LSTNet network to extract data time sequence characteristics and element correlation, and is improved for an application scene to obtain an ILSTNet network, so that the ILSTNet network extracts forward and reverse meteorological data characteristics, and also pays attention to important feature and important time step information, and the ILSTNet network is used as a first prediction network; a second prediction network is constructed by taking an encoding and decoding framework as a main body, the first prediction network and the second prediction network are dynamically weighted and fused, and a two-stage fusion prediction model MFC-IL-EDCT is obtained. The application has higher fitting accuracy for the change trend of actual meteorological data sequences and stronger data dynamic tracking capability, and therefore can better perform meteorological data authenticity verification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of meteorological data analysis, and particularly to a multi-source same-type meteorological element fusion verification method. BACKGROUND

[0002] Multi-source data fusion, also known as multi-sensor information fusion, is a comprehensive science involving computer science, network communication technology, microelectronic technology, and control technology, aiming to fuse and integrate multiple data types from different data sources to generate more comprehensive, accurate, and reliable information. In the era of big data, there are multiple data sources in massive data, and multi-source heterogeneity is one of the characteristics of big data. The dispersion of multi-source heterogeneous data divides the data into isolated islands, making data research and analysis extremely difficult. The key to improving this problem is multi-source data fusion.

[0003] According to the different levels of abstraction in the sensor data processing layer, multi-source data fusion is divided into three levels: signal-level fusion, feature-level fusion, and decision-level fusion. In the field of meteorological data, due to the influence of multiple factors on data and the large amount and diverse formats of data, feature-level and decision-level fusion are more suitable methods.

[0004] In terms of meteorological data authenticity verification, the sources of meteorological element-related data are relatively extensive, including temperature, rainfall, radiation, humidity, wind speed, topographic features, hydrogeological conditions, etc. Although these factors are independent on the surface, they are actually interconnected, mutually restrictive, and mutually related through the exchange of matter and energy. For example, during meteorological natural disasters, certain characteristics of these factors may exhibit similar abnormalities. By utilizing this correlation, we can select effective feature information for fusion to better verify authenticity.

[0005] In addition, the fusion model is a means of multi-source data fusion. The disadvantage of using a single model is that its performance is limited by data quality and the complexity of the model itself. If the data quality is insufficient or the model is too simple, it may lead to insufficient predictive ability of the model. The fusion model can utilize the advantages of multiple models to achieve better prediction or classification results, improving overall performance and generalization ability. Therefore, by analyzing, selecting, and fusing effective algorithm models, the accuracy and scientificity of meteorological data authenticity verification can be improved. This is also the advantage of feature-level fusion and decision-level fusion methods in multi-source data fusion technology.

[0006] Multi-source refers to multi-station multi-element data. Multi-station multi-element meteorological data is different from general single-station meteorological data. Affected by the diffusion flow of the atmosphere, the meteorological data of a certain meteorological station will be affected by other stations, showing the complexity of spatial and temporal changes, so the spatial correlation of multi-station meteorological data is more obvious, such as the data of the same element of different stations are often related.

[0007] For multi-station multi-element, the spatio-temporal variation characteristics of meteorological data need to be considered, that is, more attention needs to be paid to the correlation and influence between multi-station in model design, therefore, the existing processing single-station meteorological data prediction model is not applicable to multi-station multi-element situation. SUMMARY

[0008] In view of the above problems, the purpose of the present application is to provide an improved LSTNet network based on multi-feature fusion and a two-stage fusion prediction model MFC-IL-EDCT of encoder-decoder hybrid, that is, a multi-source same type meteorological element fusion verification method, which has higher fitting accuracy for the change trend of actual meteorological data sequence and stronger data dynamic tracking capability, so it can better verify the authenticity of meteorological data. The technical scheme is as follows:

[0009] A multi-source same type meteorological element fusion verification method, comprising the following steps:

[0010] Step 1: find the same type station of the meteorological observation station by using spatial correlation analysis, and introduce the target element features of the same type station; use element correlation analysis to screen the meteorological factors affecting the target element, and obtain strong correlation meteorological element features; fuse the strong correlation meteorological element features with the historical wind target element data of the target observation station to obtain a multi-feature fusion data set as the input of the model;

[0011] Step 2: introduce LSTNet network to extract data time sequence features and element correlation, and improve LSTNet network according to application scenarios to obtain ILSTNet network, so that it extracts forward and reverse meteorological data features while paying attention to important features and important time steps, and uses ILSTNet network as the first prediction network;

[0012] The second prediction network is constructed based on the coding and decoding framework, the encoder of the second prediction network selects ConvLSTM network, and the decoder part selects time convolution network;

[0013] Step 3: dynamically weighted fusion of the first prediction network and the second prediction network to obtain an improved LSTNet network based on multi-feature fusion and a two-stage fusion prediction model MFC-IL-EDCT of encoder-decoder hybrid.

[0014] Further, the step 1 specifically comprises:

[0015] Step 1.1: The minimum daily temperature is selected for analysis, combined with the latitude and longitude information between stations, and the historical data correlation coefficient, so as to establish the similarity of the underlying surface and the comparability of the same type of station meeting the requirements;

[0016] Step 1.2: Conventional meteorological elements are introduced as input features, and gray correlation analysis is introduced to screen the input feature elements, and the meteorological factors with a correlation degree higher than the set threshold are selected as the final input features;

[0017] Step 1.3: From the perspective of element correlation and spatial correlation, strong correlation meteorological element features and same type station target element features are introduced, and combined with the historical data features of the target observation station target elements, a new data set after multi-feature fusion is obtained.

[0018] Further, the step 1.1 specifically comprises:

[0019] Step 1.1.1: The spatial distance of each station is calculated using the latitude and longitude information:

[0020] S = R * arcos (cos (Y1) * cos (Y2) * cos (X1-X2) + sin (Y1) * sin (Y2))

[0021] Where (X1, Y1) and (X2, Y2) are the latitude and longitude of two meteorological observation stations, and R is the radius of the earth;

[0022] Step 1.1.2: Take the historical target element data of each station in a certain period of time for gray correlation analysis;

[0023] Step 1.1.3: Select the rule as the distance S between large monitoring stations and large monitoring stations is less than 80km, the S between large monitoring stations and regional automatic stations is less than 50km, the distance between regional automatic stations is less than 30km; The altitude difference is less than 20m, the daily minimum temperature difference is less than 2℃, and the historical target element data correlation coefficient is greater than 0.8. The station is used as a same type station.

[0024] Further, the improvement of the LSTNet network in the step 2 specifically comprises:

[0025] 1) In the cycle module and the cycle jump module, the bidirectional recurrent neural network BiLSTM is used instead of the GRU in the original LSTNet, so as to extract the data information in both directions at the same time; The update of the unit state of the cycle module and the cycle jump module at t time is expressed as follows:

[0026] f t =σ(W fx X t +W fh h t-q +bf )

[0027] i t = σ(W ix X t +W ih h t-q +b i )

[0028] s t = f t ·s t-q +i t ·g t

[0029] g t = tanh(W gx X t +W gh h t-q +b g )

[0030] o t = σ(W ox X t +W oh h t-q +b o )

[0031] h t = o t tanh(s t )

[0032] wherein q is the number of skip hidden layers, f t , i t , g t , s t , o t and h t are the forget gate, the input gate, the input node, the memory cell, the output gate and the output of the hidden layer respectively; W fx is the weight of the forget gate multiplied by the input, W ix is the weight of the input gate multiplied by the input, W gx is the weight of the input node multiplied by the input, W ox is the weight of the output gate multiplied by the input; W fh is the weight of the forget gate multiplied by the hidden layer, W ih is the weight of the input gate multiplied by the hidden layer, W gh is the weight of the input node multiplied by the hidden layer, W oh is the weight of the output gate multiplied by the hidden layer; b f , b i , b g and b oBias vectors of forget gate, input gate, input node and output gate, respectively; "·" represents the multiplication of the elements of the vectors according to the corresponding positions; σ represents the Sigmoid function; X t is an input vector; h t-q is the output of the hidden layer in the recurrent skip-connection module; s t-q is the memory cell in the recurrent skip-connection module;

[0033] Then the outputs of the recurrent and recurrent skip-connection modules are integrated by using a fully connected layer, and thus the nonlinear part ends.

[0034] 2) A channel attention mechanism CA is introduced in the convolution module, different channels are given respective weights, and the results of the convolution layer are obtained by weighting and summing all channel information.

[0035] 3) A time sequence attention mechanism TA is embedded on the side of the bidirectional recurrent neural network BiLSTM.

[0036] Further, the second prediction network first uses an encoder to mine the spatiotemporal characteristics of the data in the input feature sequence, and encodes them into a vector with a fixed dimension representation; then the decoder reads the encoded vector and decodes it to generate the predicted value of the target element at the future time;

[0037] The encoder is composed of three layers of ConvLSTM, in the three-layer ConvLSTM stacking structure, the output of the previous layer of ConvLSTM will be used as the input of the next layer of ConvLSTM; after recursively applying the ConvLSTM unit to the input sequence, the latest hidden state and the cell state will be obtained.

[0038] The decoder includes three TCN layers and a convolution layer, and also includes a flattening layer and a fully connected layer for adapting the output dimension of the target prediction value.

[0039] The final output of the encoder is used as the initial input of the decoder, the final hidden states of the three TCN layers in the decoder are spliced, and the splicing result is used as the input of the single convolution layer. After the convolution layer, the flattening layer is used for flattening and dimension reduction, and then the fully connected layer Dense is used for output. The output result is the prediction result of the target element at the next time.

[0040] Further, in step 3, assuming that the prediction value of the first prediction network at t+1 time is U(t+1), and the prediction value of the second prediction network at t+1 time is V(t+1), the final prediction result Z(t+1) of the two-stage fusion prediction model MFC-IL-EDCT is represented as follows:

[0041] Z(t+1) = w1U(t+1) + w2V(t+1)

[0042] In the formula, w1 and w2 are the weight values when the combination is optimal; under the condition of w1+w2=1, the w1 and w2 that make the mean square error (MSE) of the MFC-ILSTNet-TCN model minimum are defined as the optimal weight combination.

[0043] The beneficial effects of the present application are: the multi-feature fusion adopted in the present application can improve the prediction accuracy of the model, and has certain model adaptability; the proposed two-stage fusion prediction model MFC-IL-EDCT is beneficial to the mining of the spatio-temporal characteristics and element correlation of meteorological data, has higher fitting accuracy for the change trend of actual meteorological data sequence, has stronger data dynamic tracking capability, can realize accurate prediction of conventional meteorological elements, thereby filling in the missing values of meteorological data, and can also detect abnormal values in the observation data, realize more accurate prediction and more comprehensive detection of target elements. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 It is a structure diagram of a convolution layer with attention.

[0045] Figure 2 It is a structure diagram of BiLSTM with attention.

[0046] Figure 3 It is a structure diagram of improved LSTNet.

[0047] Figure 4 It is a structure diagram of Encoder-Decoder-ConvLSTM-TCN network.

[0048] Figure 5 It is a structure diagram of the two-stage fusion prediction model MFC-IL-EDCT of the present application.

[0049] Figure 6 It is the correlation of each site.

[0050] Figure 7 It is a correlation diagram of each meteorological element.

[0051] Figure 8 It is a schematic diagram of 10-minute average wind speed missing data.

[0052] Figure 9 It is a fitting of 10-minute average wind speed missing values.

[0053] Figure 10 It is a schematic diagram of 10-minute average wind speed abnormal data.

[0054] Figure 11 It is a schematic diagram of 10-minute average wind speed abnormal value detection. DETAILED DESCRIPTION

[0055] The application will be described in further detail below in connection with the accompanying drawings and specific embodiments.

[0056] 1. The application proposes a two-stage fusion prediction model MFC-IL-EDCT based on improved LSTNet network and encoder-decoder hybrid fusion of multiple feature fusion, that is, a multi-source same type meteorological element fusion verification method.

[0057] Firstly, the spatial correlation analysis is used to find the same type station of the observation station, and the target element characteristics of the same type station, that is, the spatial characteristics, are introduced; then the element correlation analysis is used to screen the meteorological factors affecting the target element, and the strong correlation meteorological element characteristics are obtained; the above characteristics and the historical wind target element data of the target observation station are fused to obtain MFC, so as to enrich the model characteristics and improve the predictability of meteorological data.

[0058] Secondly, the LSTNet network is introduced to extract the data time sequence characteristics and element correlation, and the LSTNet network is improved for the application scenario of the application, so that it can extract the forward and reverse meteorological data characteristics at the same time, and pay more attention to the information of important characteristics and important time steps, and the improved LSTNet network is used as one of the prediction networks.

[0059] In order to better extract the spatial characteristics of meteorological data and enhance the element applicability of the model, another prediction network is constructed based on the encoding and decoding framework, the encoder of the network selects the ConvLSTM network to ensure that the temporal relationship can be extracted while the spatial characteristics hidden in the data sequence can also be obtained; the decoder part selects a time convolutional network (temporal convolutional network, TCN) completely different from the RNN framework to decode and predict, in order to improve the accuracy of prediction.

[0060] Finally, the two prediction networks are dynamically weighted and fused to obtain the two-stage fusion prediction model MFC-IL-EDCT based on improved LSTNet network and encoder-decoder hybrid fusion of multiple feature fusion.

[0061] The experimental analysis results show that the multiple feature fusion can improve the prediction accuracy of the model and has certain model adaptability; the two-stage fusion prediction model MFC-IL-EDCT proposed has the advantages of mining the spatio-temporal characteristics and element correlation of meteorological data, higher fitting accuracy of the change trend of actual meteorological data sequence, and stronger data dynamic tracking capability, so it can better verify the authenticity of meteorological data.

[0062] 2. Two-stage fusion prediction model MFC-IL-EDCT

[0063] 2.1, Multi-feature fusion

[0064] (1) Homogeneous station wind speed feature

[0065] The spatial consistency theory is a commonly used method in meteorological data analysis, and the commonly used methods include climate statistical comparison method, spatial regression test method, and spatial interpolation method. These methods utilize the uniformity and continuity of the natural distribution of meteorological elements, and the characteristics that the meteorological data of an observation station is closely related to the data of surrounding stations. Based on the above theory, in order to enrich the model input features, the application considers finding the homogeneous station of the target observation station and introducing the target element data sequence thereof, that is, introducing the spatial feature, so as to improve the predictability of the target element data.

[0066] The commonly used method for finding reference stations is the four-orientation and eight-orientation method. The application improves the method and introduces a homogeneous model to find the homogeneous station. The homogeneous refers to the similar underlying surface. The underlying surface is the ground surface environment around the meteorological station. The traditional authenticity test method mostly takes the Euclidean distance as the standard when selecting reference stations or similar stations, which causes the authenticity test control result to be greatly affected by the topography. According to the knowledge of weather dynamics and atmospheric thermodynamics, the underlying surface is closely related to the local meteorological elements, so the application selects the daily minimum temperature for analysis, and further combines the longitude and latitude information between stations, the altitude information, and the historical data correlation coefficient, so as to establish the homogeneous station with similar underlying surface and high comparability. The specific process is as follows:

[0067] (a) The spatial distance between stations is calculated by using the longitude and latitude information of each station. Assuming that the longitude, latitude and altitude of two meteorological observation stations are (X1, Y1, H1) and (X2, Y2, H2) respectively, and the radius of the earth is R, then the spatial distance S between the two stations is calculated according to the following formula:

[0068] S = R * arcos (cos (Y1) * cos (Y2) * cos (X1-X2) + sin (Y1) * sin (Y2))

[0069] (b) The 3-year historical target element data of each station is taken for grey correlation analysis. The grey correlation analysis method is not affected by the sample size, and is not limited by whether the sample has a rule or not, and is very suitable for calculating the correlation degree between multiple time series.

[0070] (c) The selection rule of the same type station is: the selection rule is that the distance S between the large monitoring station and the large monitoring station is less than 80km, the S between the large monitoring station and the regional automatic station is less than 50km, the distance between the regional automatic stations is less than 30km; the altitude difference is less than 20m, the daily minimum temperature difference is less than 2℃, and the historical target element data correlation coefficient is greater than 0.8. Thus, the same type station with high correlation and good data comparability of the target observation station is obtained, and the target element data thereof is taken as one of the input features.

[0071] (2) Strongly correlated weather element features

[0072] Meteorological data prediction is a spatiotemporal prediction problem, and the prediction value is related to not only its own historical data but also other meteorological elements. Therefore, conventional meteorological elements can be considered as input features. However, if all these meteorological factors are introduced, the complexity of the network will undoubtedly increase, and the feature dimension disaster will be caused. Therefore, the grey correlation analysis is also introduced to screen the above-mentioned feature elements, and the meteorological factors with high correlation degree are selected as input features, so as to enrich the features, avoid information redundancy, and improve the data predictability.

[0073] (3) Multi-feature fusion MFC

[0074] In order to enrich the model input features and improve the data predictability, the present application introduces the strongly correlated weather element features and the same type station target element features from the element correlation and spatial correlation angles, and combines the target observation station target element historical data features to obtain a new data set after multi-feature fusion. When the three features are fused, the time corresponding method is adopted. For example, when the wind speed S(t+1) at t+1 moment needs to be predicted, the data of the above-mentioned three features at the previous n moments are taken as the input, and the corresponding input feature vector X(t) is as follows.

[0075]

[0076] Among them, S1(t) to S n (t) represents the wind speed sequence of n same type stations, W1(t) to W n (t) represents the screened strongly correlated weather element sequence, S'(t) is the historical wind speed sequence of the target station. N represents the width of the set time window. Each row represents different feature element data at the same moment, and each column represents the data of the same feature element at different moments within the time window (t-n+1, t). Thus, the three feature fusions are completed, and a new data set is formed.

[0077] 2.2, LSTNet network

[0078] The LSTNet network is composed of linear and nonlinear parts, and the specific Figure 1The nonlinear part includes a convolution module, a recurrent module, and a special recurrent skip module, and the linear part is composed of an autoregressive module. Such a combination can successfully capture long-term and short-term patterns in the data, making the model more robust.

[0079] In the nonlinear part, the multivariate time series is used as the input of the convolution module, which adopts a one-dimensional convolution without a traditional pooling layer to capture the dependency between multiple variables and the short-term local features of each variable in the time dimension. The following formula represents the convolution of the nth filter on the variable matrix X.

[0080] h n = RELU(W n *X+b n )

[0081] where the feature vector output after the convolution operation is h n , RELU is an activation function, W n represents the weight matrix of the convolution kernel, W n is connected to the nth feature map, X is the input vector of the convolution module, and b n is the bias. Assuming there are i filters, the convolution module outputs a feature vector of size i*j, and j is the length of the vector generated after the first convolution kernel is applied to X.

[0082] The data output by the convolution layer will be simultaneously input into the recurrent module and the recurrent skip module. The present application uses LSTM as the basic unit in both the recurrent module and the recurrent skip module, and the update of the unit state at time t can be expressed as follows:

[0083] f t = σ(W fx X t +W fh h t-q +b f )

[0084] i t = σ(W ix X t +W ih h t-q +b i )

[0085] s t = f t ·s t-q +i t ·gt

[0086] g t = tanh(W gx X t +Wgh h t-q +b g )

[0087] o t =σ(W ox X t +W oh h t-q +b o )

[0088] h t =o t tanh(S t )

[0089] where q is the number of skip-connection hidden layers, f t , i t , g t , s t , o t and h t are the forget gate, input gate, input node, memory cell, output gate and hidden layer output respectively; W fx is the weight of the forget gate multiplied by the input, W ix is the weight of the input gate multiplied by the input, W gx is the weight of the input node multiplied by the input, W ox is the weight of the output gate multiplied by the input; W fh is the weight of the forget gate multiplied by the hidden layer, W ih is the weight of the input gate multiplied by the hidden layer, W gh is the weight of the input node multiplied by the hidden layer, W oh is the weight of the output gate multiplied by the hidden layer; b f , b i , b g and b o are the bias vectors of the forget gate, input gate, input node and output gate respectively; "·" represents the multiplication of the elements of the vector according to the corresponding position; σ represents the Sigmoid function; X t is the input vector; h t-q is the hidden layer output in the recurrent skip-connection module; s t-q is the memory cell in the recurrent skip-connection module.

[0090] The linear part is added by introducing an autoregressive AR to add a linear part for prediction. When some data does not have strong periodicity, only the nonlinear part is used for prediction, and the effect is not ideal. It is found that after adding the linear prediction part, the low prediction accuracy of the neural network model in non-periodic changes can be made up. The formula of the autoregressive part is as follows:

[0091]

[0092] where, represents the prediction result of the autoregressive part, and b ar is the coefficient of the model, q ar is the window size of the input matrix. The final prediction result contains a linear part and a nonlinear part The formula is as follows:

[0093]

[0094] It is known that meteorological data is also a multivariate time series, so the application selects LSTNet which performs well on multivariate time series as one of the prediction networks. The first layer structure of LSTNet is a convolutional neural network without a pooling layer, which can extract short-term features in the time dimension of meteorological data and extract local dependencies between variables; and the special recurrent skip module utilizes the periodicity of the input time series signal, so that it is possible to capture very long-term dependencies; in addition, the AR autoregressive process adds a linear component to the prediction, so that the output can respond to the scale change of the input. In summary, the application selects LSTNet as one of the prediction networks. However, LSTNet also has some shortcomings for the application scenario of the application, so it will be improved in the following to make it more suitable for the application scenario of the application in order to achieve better prediction effect.

[0095] 2.3, improved LSTNet network

[0096] The application selects LSTNet as one of the prediction networks, and improves LSTNet for the accuracy of meteorological data prediction as follows:

[0097] (1) In the cycle module and the cycle jump module, the original LSTNet uses GRU as the basic unit, only considers single-direction data and ignores the influence of reverse sequence information on prediction. For this, the application proposes to replace GRU with bidirectional recurrent neural network BiLSTM, which facilitates the extraction of bidirectional data information and obtains better prediction effect.

[0098] (2) The channel attention mechanism is introduced in the convolution module. When the traditional convolution layer extracts information, the weights of each channel of the feature map are allocated equally, but in fact, the important degree of information contained in different channels is different, which causes the loss of important information. In order to strengthen the learning of important channel features by the model, the model proposed in the application adds an attention mechanism in the CNN, assigns respective weights to different channels to form a channel attention mechanism CA, and finally weights all channel information to obtain the result of the convolution layer, as shown in Figure 1 .

[0099] The timing attention mechanism is introduced in the cycle module and the cycle jump module. Whether the GRU or the BiLSTM is used as a basic unit, information loss occurs in the memory of a long time sequence. In order to highlight the information of important time steps when the time sequence dependency is mined, the BiLSTM side is embedded with a timing attention mechanism TA, and the structure diagram of the attention mechanism is as shown in Figure 2 .

[0100] Yt represents the tth feature vector output by the BiLSTM, Y t The initial state vector Z t is obtained after passing through the attention mechanism layer, and the initial state vector is multiplied by the weighting coefficient to obtain the final output vector O. In summary, the improved LSTNet structure is as shown in Figure 3 .

[0101] 2.4, Encoder-Decoder-ConvLSTM-TCN network

[0102] Another prediction network (encoder-decoder-convlstm-tcn, EDCT) proposed in the application adopts an encoder-decoder framework widely used in the fields of automatic text summarization and spatiotemporal sequence prediction in deep learning, so as to obtain better prediction performance. The whole model is divided into two parts of an encoding network and a prediction network. The encoding part takes ConvLSTM as a basic unit, and the decoding part is composed of a TCN network. First, the encoder is used to mine the spatiotemporal characteristics of the data in the feature sequence of the input, and encode into a fixed-dimension represented vector; then the decoder reads the encoded vector and decodes it to generate the prediction value of the target element at the future time. The specific structure of the Encoder-Decoder-ConvLSTM-TCN network is as shown in Figure 4 .

[0103] Figure 4 The left half part is the encoder, and the bottom is the input X of the model X = {X1, X2, X3…X m}, m is the number of features selected by the MFC, and if the input time window is n, then the m features at the first n time points constitute the matrix in the figure. The encoding part is composed of three layers of ConvLSTM, and the ConvLSTM unit is used because it can extract the time correlation of the input sequence, and at the same time, the convolution operation can mine the spatial correlation.

[0104] In the three-layer ConvLSTM stacking structure, the output of the previous layer of ConvLSTM will be used as the input of the next layer of ConvLSTM, and the convolution operation can determine the future cell state and whether to activate the three gates, after applying the ConvLSTM unit to the input sequence recursively, the latest hidden state can be obtained and cell state Both will be used for the decoder on the right half, that is, the prediction network. The input, output, state vector and gating unit of each layer of ConvLSTM are three-dimensional tensors, compared with the two-dimensional vector of traditional FC-LSTM (although the input, output and hidden state of FC-LSTM can also be regarded as a three-dimensional tensor with the last two dimensions being 1), the additional one dimension is the spatial correlation information extracted by ConvLSTM, so ConvLSTM better fits the characteristics of spatio-temporal sequence problems.

[0105] Figure 4 The right half of the prediction network is the decoder, that is, the prediction network. The structure of the prediction network is similar to that of the encoding network, including three TCN layers and a convolutional layer, the only difference is that there is a flattening layer and a fully connected layer to adapt to the output dimension of the target prediction value.

[0106] The hidden state and cell output value in the initial state of the prediction network are determined by the last state of the encoding network, that is, the final output of the encoding network is used as the initial input of the prediction network, which is also the connection method of the encoder and the decoder. On this basis, the final hidden state of the three TCN layers in the prediction network is spliced, and the splicing result is used as the input of the single convolutional layer, and the convolutional layer here is a single 1*1 convolutional layer, so as to ensure that the output and input of the convolutional layer have the same latitude, that is, three-dimensional vectors; After the convolutional layer, the Flatten layer is used for flattening and dimension reduction, and then the Dense layer is used for output, and the output result is the prediction result of the target element at the next time.

[0107] Common decoders include RNN, LSTM, GRU, etc. The present application uses TCN as the decoding network. At present, various variants of RNN have achieved good results in time series prediction, but outside the RNN framework, many studies have found that in tasks such as machine translation and speech recognition, time convolution network TCN can defeat RNN and achieve better results. TCN is a neural network model that uses convolutional neural networks for time series modeling. Its own dilated convolution and residual structure make it avoid the shortcomings of RNN, have stable gradient, can also process feature sequences in parallel, train faster, and have fewer parameters when building the model, which is a completely different structure from RNN. Therefore, the present application selects TCN as the decoding network.

[0108] 2.5, two-stage fusion prediction model MFC-IL-EDCT

[0109] In order to solve the problem of single data characteristics in meteorological data prediction, the present application proposes a multi-feature fusion method based on multi-station multi-element data, which supplements the input features from the aspects of spatial correlation and element correlation. In addition, in order to fully extract the spatial and temporal information contained in the input data sequence, enhance the adaptability of the model to meteorological elements, and improve the overall prediction accuracy, the present application proposes a two-stage fusion prediction model MFC-IL-EDCT. First, the LSTNet network is improved to obtain the ILSTNet network, which is used as one of the prediction networks. Then, the ConvLSTM network and the TCN network are used to construct another prediction network Encoder-Decoder-ConvLSTM-TCN with the encoder-decoder as the main body. Finally, the two networks are dynamically weighted and fused to obtain the MFC-IL-EDCT model, as shown in the structure of Figure 5

[0110] Figure 5 In the above formula, the original time series is preprocessed according to certain rules, and the spatial correlation and element correlation are analyzed to screen out the target element features of the same type station and other meteorological element features with high correlation, and then combined with the historical target element sequence of the target observation station to form a combined feature sequence X(t) as the input of the model.

[0111] Then, the ILSTNet network and the Encoder-Decoder-ConvLSTM-TCN network are used to predict the hourly ground observation meteorological data, i.e. to predict the target element value at the next time point based on the values of each feature element at the previous n time points. The optimizer in the training process of the two networks is Adam, the maximum number of iterations is set to 500, and the checkpoint is set to save the optimal model parameters, and then the optimal model is tested. Assuming that the prediction value of the ILSTNet model at t+1 time is U(t+1), and the prediction value of the Encoder-Decoder-ConvLSTM-TCN network at t+1 time is V(t+1), then the final prediction result Z(t+1) of the MFC-IL-EDCT model can be represented as follows.

[0112] Z(t+1)=w1U(t+1)+w2V(t+1)#(4-12)

[0113] In the above formula, w1 and w2 are the optimal combination weights. The key of the fusion model is to determine the weights, and under the condition of w1+w2=1, the w1 and w2 that make the mean square error MSE of the MFC-ILSTNet-TCN model minimum are defined as the optimal weight combination.

[0114] 3. Experimental results and analysis

[0115] ​During the experiment, firstly, similar stations were selected according to certain rules and multi-feature fusion was performed; secondly, the two-stage fusion prediction model MFC-IL-EDCT was analyzed in a prediction experiment; finally, the model was applied to the authenticity verification of meteorological observation data, that is, to realize the processing of missing data and the detection of abnormal data.

[0116] 3.1 Multi-feature fusion results

[0117] The MFC consists of three parts: target element characteristics of similar stations, strongly correlated meteorological element characteristics, and historical target element characteristics of the target observation station. The previous text has explained how to select target element characteristics of similar stations and strongly correlated meteorological element characteristics. The following text shows the screening results and then constructs the MFC.

[0118] (1) Characteristics of target elements of similar stations

[0119] Taking the Fuqing meteorological observation station (area station number 58942) in Fujian Province as an example, the 10-minute average wind speed of this station is used as the target predicted value. Table 1 shows some meteorological observation stations around this station that meet the above-mentioned distance, altitude, and daily minimum temperature difference conditions.

[0120] Table 1. Location information of surrounding meteorological stations

[0121]

[0122] Grey relational analysis was performed on the 10-minute average wind speed sequences of these observation stations from 2016 to 2018 and the 10-minute average wind speed data of the target observation station, Fuqing. The results... Figure 6 As shown.

[0123] Figure 6 The numbers 0 to 6 represent the automatic wind speed monitoring stations in Fuqing, Minhou, Lianjiang, Yongtai, Changle, and Mawei, respectively. It can be seen that the 10-minute average wind speed data of each station, after being filtered based on distance, altitude, and daily minimum temperature difference, shows strong correlation. To ensure prediction accuracy while reducing network complexity, a correlation of 0.85 was used as the filtering threshold. The final selected stations are: Minhou, Yongtai, and Mawei Tianma Park. The 10-minute average wind speed data from these three stations were chosen as the target feature for these stations.

[0124] (2) Characteristics of strongly correlated meteorological elements

[0125] Meteorological elements are interrelated. Grey relational analysis of conventional meteorological elements and 10-minute average wind speed can identify meteorological factors that significantly influence wind speed, thus improving the accuracy of forecasts. The results of the grey relational analysis are as follows: Figure 7 As shown.

[0126] Figure 7 0 to 8 in the above table represent 10-minute average wind speed, sea level pressure, air temperature, relative humidity, hourly precipitation, 10-minute average wind direction, 2-minute wind direction, 2-minute average wind speed, 1-meter visibility in turn. It can be seen that the meteorological factors with a correlation greater than 0.85 with the 10-minute average wind speed are the 2-minute average wind speed and the hourly precipitation, so the two are selected as the strongly correlated meteorological element features.

[0127] (3) MFC construction

[0128] After the above feature selection, the input feature vector of the model can be determined as follows:

[0129]

[0130] In the formula, S1(t), S2(t), S3(t) represent the wind speed data of the three same type stations, W1(t) represents the 2-minute average wind speed sequence of the target observation station, W2(t) represents the hourly precipitation sequence of the target observation station, and S'(t) is the historical 10-minute average wind speed of the target observation station itself. n represents the width of the set time window. Each row represents different feature element data at the same time, and each column represents data of the same feature element at different times within the time window (t-n+1, t).

[0131] 3.2, the realness test experiment of the MFC-IL-EDCT model of the application on meteorological observation data

[0132] Various errors may be generated in the process from collection to storage of meteorological observation data. Here, mainly aiming at two types of error types of data missing and data abnormality, the following experiments are done.

[0133] (1) Missing value fitting

[0134] Taking the 10-minute average wind speed as an example, the data of the new ground meteorological automatic observation station in Fujian Province is taken for research, and it is found that the wind speed data has missing phenomenon from 10 o'clock on May 11, 2019 to 20 o'clock on May 11, as shown in Figure 8 . And at this time, other element data with high correlation with the 10-minute average wind speed are relatively complete. Therefore, the two-stage fusion prediction model MFC-IL-EDCT proposed by the application can be used to predict the missing values of wind speed according to other element data and historical wind speed data. The fitting of the missing values of the 10-minute average wind speed is shown in Figure 9 .

[0135] (2) Abnormal value detection

[0136] For the data of October 2019, it is found that the temperature observation value from 12:00 on October 18 to 19:00 on October 18 is abnormal, as shown in Figure 8 Normally, the 10-minute average wind speed value of this period should be about 2 to 9, but the recorded observation value is between 15 to 18, which is obviously abnormal and needs to be verified.

[0137] For the data of October 2019, it is found that the temperature observation value from 12:00 on October 18 to 19:00 on October 18 is abnormal, as shown in Figure 10 Normally, the 10-minute average wind speed value of this period should be about 2 to 9, but the recorded observation value is between 15 to 18, which is obviously abnormal and needs to be verified. Figure 11

[0138] As can be seen from Figures 8 to 11 It can be seen that the two-stage fusion prediction model MFC-IL-EDCT proposed by the present application can realize accurate prediction of conventional meteorological elements, thereby filling in missing meteorological data, and at the same time can detect abnormal values in observation data, realizing more accurate prediction and more comprehensive detection of target elements.​

Claims

1. A multi-source homogeneous meteorological element fusion verification method, characterized in that, The method comprises the following steps: Step 1: find the homotype station of the meteorological observation station by using spatial correlation analysis, and introduce the target element characteristics of the homotype station; Step 2: introduce the LSTNet network to extract data time sequence characteristics and element correlation, and improve the LSTNet network according to the application scene to obtain the ILSTNet network, so that the ILSTNet network extracts the forward and reverse meteorological data characteristics while paying attention to important characteristics and important time step information, and the ILSTNet network is used as a first prediction network; Step 3: dynamically weighted fusion of the first prediction network and the second prediction network to obtain the improved LSTNet network based on multi-feature fusion and the dual-stage fusion prediction model MFC-IL-EDCT of the encoder-decoder hybrid; In step 2, the LSTNet network comprises linear and nonlinear parts, the linear part is composed of an autoregressive module, and the nonlinear part comprises a convolution module, a recurrent module and a recurrent skip module; The improvement of the LSTNet network specifically comprises: Then the outputs of the recurrent module and the recurrent skip module are integrated by using a full connection layer, and the nonlinear part ends; 2) introducing a channel attention mechanism in the convolution module, giving different channels their own weights to form a channel attention mechanism CA, and weighting and summing all channel information to obtain the result of the convolution layer; 1) In the recurrent module and the recurrent skip module, the BiLSTM is used instead of the GRU in the original LSTNet, so as to extract the bidirectional data information at the same time; the recurrent module and the recurrent skip module update the state of the time unit as follows: The update of the state of the time unit is as follows: ； ； ； ； ； ； wherein, is the number of skip-connection hidden layers, , , , , and are the forget gate, input gate, input node, memory cell, output gate and output of the hidden layer, respectively; is the weight multiplied by the input, is the weight multiplied by the input, is the weight multiplied by the input node, is the weight multiplied by the output gate; is the weight multiplied by the hidden layer, is the weight multiplied by the input gate and hidden layer, is the weight multiplied by the input node and hidden layer, is the weight multiplied by the output gate and hidden layer; , , and are the bias vectors of the forget gate, input gate, input node and output gate, respectively; ” means that the elements of the vector are multiplied according to the corresponding positions; represents the Sigmoid function; is the input vector; is the output of the hidden layer in the recurrent skip-connection module; is the memory cell in the recurrent skip-connection module; 3) embedding a time sequence attention mechanism TA on the side of the bidirectional recurrent neural network BiLSTM. Step 1 specifically comprises: Step 1.1: selecting daily minimum temperature for analysis, and combining latitude, longitude and elevation information of the stations and historical data correlation coefficients to establish homotype stations with satisfactory underlying surface similarity and comparability; 2. The multi-source homogeneous weather element fusion verification method according to claim 1, characterized in that, Step 1.2: introducing conventional meteorological elements as input characteristics, and introducing gray correlation analysis to screen the input characteristic elements, and selecting meteorological factors with a correlation degree higher than a set threshold as the final input characteristics; Step 1.3: introducing strong correlation meteorological element characteristics and homotype station target element characteristics from the aspects of element correlation and spatial correlation, and combining target observation station target element historical data characteristics to obtain a new data set after multi-feature fusion. Step 1.1 specifically comprises: Step 1.1.1: calculating the spatial distance of each station by using the latitude and longitude information of each station:

3. The multi-source homogeneous weather element fusion verification method according to claim 2, characterized in that, Step 1.1.2: taking the historical target element data of each station in a specific time period for gray correlation analysis; The second prediction network first uses an encoder to mine the spatial and temporal characteristics of the data in the input feature sequence, and encodes it into a fixed-dimensional representation vector; then the decoder reads the encoded vector and decodes it to generate a predicted value of the target element at a future time; ； wherein, and are the longitude and latitude of the two weather observation stations, respectively, is the radius of the earth; ​ Step 1.1.3: Select the rule as the distance between large stations and large stations less than 80 km, the distance between large stations and regional automatic stations less than 50 km, the distance between regional automatic stations is less than 30 km; the elevation difference is less than 20 m, the daily minimum temperature difference is less than 2℃, and the historical target element data correlation coefficient is greater than 0.

8. The site is selected as a same type site.

4. The multi-source homogeneous weather element fusion verification method according to claim 1, characterized in that, ​ The encoder is composed of three layers of ConvLSTM, and the output of the previous layer of ConvLSTM will be used as the input of the next layer of ConvLSTM in the three-layer ConvLSTM stack structure; After recursively applying the ConvLSTM unit to the input sequence, the latest hidden state is obtained and cell state , both of which will be used in the decoder; The decoder includes three TCN layers and a convolutional layer, and also includes a flattening layer and a fully connected layer for adapting the output dimension of the target prediction value; The final output of the encoder is used as the initial input of the decoder, the final hidden states of the three TCN layers in the decoder are spliced, and the splicing result is used as the input of the single convolutional layer. After the convolutional layer, the flattening layer is used for flattening and dimension reduction, and then the fully connected layer Dense is used for output. The output result is the prediction result of the target element at the next moment.

5. The multi-source homogeneous weather element fusion verification method according to claim 1, characterized in that, The step 3 assumes that the first prediction network predicts the value at time t+1 as , and the second prediction network predicts the value at time t+1 as , then the final prediction result of the two-stage fusion prediction model MFC-IL-EDCT is represented as follows: ​ ； In the formula, and is the weight value when the combination is optimal; under the condition of satisfying , define and is the optimal weight combination.

Images