EMD-LSTM-based wind speed prediction method fusing spatial features

Abstract

The invention relates to an EMD-LSTM (empirical mode decomposition-long short-term memory)-based wind speed prediction method fused with spatial features. According to the method, time sequence and space data are acquired from a monitoring station, a data set is constructed through preprocessing, a space-time prediction model is established for wind speed prediction, two prediction results are fused through weight Bayesian optimization, and a prediction result is obtained. EMD decomposition is carried out on wind speed rate characteristics in the time model, a plurality of LSTMs are input, and prediction components of the LSTMs are accumulated to obtain a wind speed prediction value only considering time sequence characteristics. Vector decomposition and interpolation are carried out on the wind speed of adjacent stations in the space model, a wind speed vector is restored to obtain the interpolation of the wind direction and the wind speed rate, and the distance with the actual value is calculated to obtain the space independence weight. And performing weight Bayesian optimization in model fusion according to a spatial independence weight to obtain wind speed value prediction fused with spatial-temporal characteristics. Influences of related time features and space features in a wind speed prediction task are considered at the same time, and the model structure and the prediction method are more reasonable.

Description

Technical Field

[0001] The invention belongs to the field of meteorological forecasting, and specifically is a wind speed forecasting method based on EMD-LSTM fusion spatial features. Background Art

[0002] The transmission and diffusion of urban air pollutants are significantly affected by wind direction and wind speed. Wind direction determines the direction of pollutant transmission and diffusion, while wind speed directly affects the speed of transmission and diffusion. Therefore, if an accurate wind speed prediction model is constructed based on wind speed and related environmental data series, it can better understand the diffusion law of urban air pollutants, evaluate the distribution and concentration of pollutants, and provide important reference for environmental regulatory departments to deal with urban pollution problems.

[0003] In the wind speed forecasting task, the classic methods usually include basic models based on statistics and composite algorithms that integrate multiple basic models. Basic models include moving average models, autoregressive models, etc., while composite algorithms include autoregressive moving average models (ARMA), autoregressive integrated moving average models (ARIMA), etc. Although these classic models perform well in simple tasks, they are usually only applicable to single-variable time series forecasting and have difficulty capturing nonlinear relationships in data. With the rapid development of deep learning, deep recurrent neural networks represented by GRU and LSTM have attracted much attention in the field of time series forecasting due to their strong versatility and high prediction accuracy. However, in the wind speed forecasting task, the relevant spatial features will also have an important impact on the prediction results. Therefore, a single recurrent neural network model usually performs poorly in the wind speed forecasting task. Summary of the invention

[0004] In order to better support the air pollution supervision work of the ecological environment department and to conduct rapid prediction and evaluation before and after pollution occurs, the present invention provides a wind speed prediction method based on EMD-LSTM fusion spatial features, which can more accurately predict urban wind speed, thereby providing scientific protection for the air quality safety of the people.

[0005] The technical solution adopted by the present invention based on the above-mentioned purpose is as follows:

[0006] A wind speed prediction method based on EMD-LSTM and fusion spatial features includes the following steps:

[0007] Collect time series data and spatial data in the target monitoring area and mark the actual wind direction observation value D at the current moment l , the actual observed value of wind speed V t ; Fill in missing values ​​and outliers, and construct time series data sets and spatial direction components;

[0008] Perform EMD decomposition on the wind speed time series in the time series data to extract wind speed components of different time scales; combine all time series features as input, iteratively train the LSTM time series model to output the prediction component, and accumulate the wind speed prediction value in the time series space after denormalization; use the trained LSTM time series model to predict the wind direction time series prediction value D at the next moment t and wind speed time series prediction value Vl;

[0009] The interpolation of the spatial direction component is calculated, and the wind speed is vector-reduced to obtain the interpolation prediction value D of the wind direction at the current moment. p and wind speed interpolation prediction value V p ; Calculate spatial independence weight: W d_a and W v_a ;

[0010] Perform spatiotemporal Bayesian optimization on the spatial independence weights; obtain the optimized spatial independence weights W a_z , W d_z , establish a fusion prediction model to predict the wind speed vector in time and space;

[0011] Output the final wind direction prediction value D z and the final wind speed prediction value V z .

[0012] The time series data includes monitoring time, wind speed, wind direction, temperature, air pressure, and humidity; the spatial data includes site number, longitude and latitude.

[0013] The constructing of the sequence data set and the spatial component includes:

[0014] Time series data set: Generate corresponding seasonal features according to the monitoring time, perform one-hot encoding, and normalize;

[0015] Spatial component: select the k neighboring stations with the closest Euclidean distance to the target monitoring station, and further decompose the wind speed vector data of all neighboring stations at that moment into wind speed components in the east-west and north-south directions.

[0016] The wind speed sequence component characteristics at different time scales in the EMD decomposition are short-term fluctuation characteristics, periodic characteristics or local characteristics further divided into fine granularity.

[0017] The loss function of the iterative training LSTM model is defined as:

[0018] The wind speed loss is calculated using the mean square error loss function, and the wind direction loss is calculated using the cyclic loss function to process the periodicity of the angle. The overall loss is the weighted sum of the wind speed loss and the wind direction loss.

[0019] The interpolation calculation for the spatial direction component is performed by using IDW inverse distance weighted calculation.

[0020] The spatial independence weight of wind direction W a _a. Spatial independence weight of wind speed W v_a W d_a =|D t -D p |, W v_a =|V t -V p |.

[0021] is the two spatial independence weights: W d_ a and W v_a , respectively into the following mapping function W a Perform Bayesian optimization of spatiotemporal weights to obtain the optimized spatial independence weight W d_z , W d_z ;

[0022]

[0023] Among them, W a represents the independence weight term, which is used to adjust the weights of the time series prediction value and the interpolation prediction value to reflect the degree to which the target site is affected by the overall surrounding monitoring points; A m Represents the lower limit of the difference range, A n Represents the upper limit of the difference range; W m Represents the upper limit of weight allocation ratio, W n The upper limit of the proportion of representative weight distribution;

[0024] In the above formula, if we calculate W z <1, then W is not correct z Make adjustments; otherwise, if W z When >1, it means that the spatial characteristics of the surrounding monitoring points are not considered, so W is set z =1.

[0025] The fusion prediction model is:

[0026] D z =(1-W d_z )*D p +W d_z *D l

[0027] V z =(1-W v_z )*V p +W v_z *V l

[0028] Among them, D zis the final wind direction prediction value, V z is the final wind speed prediction value.

[0029] The present invention has the following beneficial effects and advantages:

[0030] 1. The algorithm is highly automated. For a specific monitoring site, the present invention only requires one-time LSTM training and spatiotemporal weight search, after which the wind speed can be automatically predicted without human intervention, thus saving a lot of labor costs.

[0031] 2. The algorithm has strong versatility. The idea of ​​the present invention has good versatility and does not require additional adaptation for scenes in different time and regions. It can be easily extended to wind speed prediction tasks in various scenes.

[0032] 3. High algorithm accuracy. The present invention decomposes and extracts features of meteorology at multiple scales, takes the LSTM model as the basic model, and weightedly integrates the spatial features between monitoring sites. This is a more reasonable new wind speed prediction method. Compared with the previous prediction method that only considers time series features, this method has higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a structural diagram of the wind speed prediction method based on EMD-LSTM fusion spatial features of the present invention; DETAILED DESCRIPTION

[0034] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation method of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the invention, so the present invention is not limited by the specific implementation disclosed below.

[0035] Step 1: Data acquisition. Environmental monitoring sites are equipped with a variety of environmental data monitoring equipment (such as wind speed and direction meters, thermometers, barometers, hygrometers, etc.). These sites store the collected data in a central database. By retrieving from this database, the time series data recorded by the target environmental monitoring site is obtained: including monitoring time, wind speed, wind direction, temperature, air pressure, humidity, and the actual observed value of the wind direction at the current moment is marked as D t The actual observed wind speed is V t ; At the same time, obtain the spatial data of all environmental monitoring sites including site number, longitude and latitude.

[0036] Step 2: Data cleaning: Preprocess the time series data and spatial data of the monitoring sites, mainly including missing value and outlier processing.

[0037] Step 3: Perform feature conversion on the time series data of the monitoring sites, including generating corresponding seasonal features according to the monitoring time and performing one-hot processing on the seasonal features; normalizing the data; and dividing the data into training and test sets.

[0038] Step 4: Screen the neighboring subsets of the target monitoring site, select the k neighboring sites with the closest Euclidean distance to the target monitoring site, and decompose the wind speed vector data of all neighboring sites at that moment into the wind speed in the east-west direction and the wind speed in the north-south direction.

[0039] Step 5: Based on the time series data obtained in step 3, decompose the wind speed time series to extract the short-term fluctuation characteristics, periodic characteristics or other local characteristics of the wind speed series.

[0040] Step 6: LSTM time series model training and testing. The decomposed multiple feature sequences and the remaining features (wind direction, temperature, air pressure, humidity) are combined and put into LSTM as input features for training. The extracted features are fully learned and the predicted values ​​are denormalized. Then, the predicted value components of each model are accumulated to obtain the prediction results, and the model is tested and verified.

[0041] Step 7: LSTM time series model prediction. Use the trained LSTM model to make predictions and obtain the wind direction time series prediction value D at the next moment in the future. l And the wind speed time series prediction value V l .

[0042] Step 8: Using the wind speed data obtained in step 4, interpolate the nearest k neighboring stations in the east-west and north-south directions respectively through IDW inverse distance weight calculation, and restore the wind speeds in the two directions by vector to obtain the interpolated wind direction prediction value D at the current moment. p and wind speed interpolation prediction value V p .

[0043] Step 9: Extract the independence weights of monitoring sites. Evaluate the predicted values ​​obtained in step 8 to obtain the spatial independence weights of wind direction and wind speed.

[0044] Step 10: Based on the interpolation prediction value obtained in step 8 and the spatial independence weight obtained in step 10, combined with the time series prediction value obtained in step 7, perform spatiotemporal weight Bayesian optimization to obtain a fusion prediction model.

[0045] Step 11: Obtain the prediction output according to the fusion prediction model optimized in step 10, and obtain the final wind direction prediction value D of the fusion spatial features. z and the final wind speed prediction value Vz .

[0046] The data cleaning specifically includes: using the k-nearest neighbor algorithm to detect outliers; for short-term missing data, forward fill, backward fill and mean fill are used simultaneously to ensure the filling coverage, and smoothing is performed.

[0047] The one-hot processing of seasonal features is specifically as follows: converting seasonal information into a numerical form that can be used by the model. Using one-hot encoding, each season is represented as a vector whose length is equal to the number of seasons, and has a 1 at the position corresponding to the season and 0 at other positions.

[0048] The normalization operation is: using min-max standardization to map the data to [0, 1].

[0049]

[0050] In the above formula, f old is the original eigenvalue, f min is the minimum value of the feature in the sample set, f max is the maximum value of the feature in the sample set.

[0051] The selection of the nearest k adjacent sites: a suitable k value should be selected according to the number of sites and the density of sites. The embodiment selects a k value of 12 through control experiments.

[0052] The decomposition of the wind speed time series is specifically as follows: using EMD to decompose the wind speed data with non-stationary and nonlinear properties into multiple relatively stable intrinsic mode functions (IMFs) and residuals (RES), each IMF represents fluctuations on different time scales. The EMD method can remove the non-stationarity of the wind speed series, extract the short-term fluctuation characteristics, periodic characteristics or other local characteristics of the wind speed series, and thus provide a more fine-grained prediction result.

[0053] The training process of the LSTM time series prediction is:

[0054] 1) Input the time step to obtain multiple prediction value components. And accumulate the prediction value components of each LSTM model to obtain the wind speed time series prediction result V considering multiple time scales l , and predict the wind direction D l .

[0055] 2) Embodiment Through continuous improvement of model parameters, it is known that selecting a time step of 20, a hidden layer size of 256, and using the AdamW method for model optimization can achieve the best prediction effect.

[0056] 3) Definition of loss function: Use mean square error loss function to calculate wind speed loss, use cyclic loss function to process the periodicity of angles, and calculate wind direction loss. The overall loss is the weighted sum of wind speed loss and wind direction loss, where the wind speed loss weight is 1.0 and the wind direction loss weight is 0.5. Training ends when the total loss converges.

[0057] The specific operations of decomposing and restoring the wind speed vector are as follows:

[0058] 1) In the decomposition stage: According to the angle value of the wind direction, the wind speed vector is decomposed into an east-west component (marked as U) and a north-south component (marked as V).

[0059] 2) In the restoration stage: the inverse tangent function is used to calculate the final wind direction angle value for the wind speeds in the two directions obtained by interpolation, and the complete wind speed information is calculated at the same time.

[0060] The specific operation of extracting the independence weight of the monitoring site is: respectively convert the wind direction interpolation prediction value D p , the wind speed interpolation prediction value Vp and the corresponding wind direction actual observation value D t The actual observed value of wind speed is V t The absolute value of the difference between them is used as an additional weight term W d_a and W v_a , which can be expressed as:

[0061] W d_a =|D t -D p |

[0062] W v_a =|V t -V p |

[0063] The specific operation of the spatiotemporal weight Bayesian optimization is as follows: according to the independence weight term W d_a and W v_a In view of the negative correlation between the target site and the wind direction / wind speed of the surrounding monitoring points, mapping functions are set respectively. The unified form of the formula is as follows:

[0064]

[0065] Among them, W a represents the independence weight term, i.e., the difference, which is used to adjust the weights of the time series prediction value and the interpolation prediction value to reflect the degree to which the target site is affected by the overall surrounding monitoring points. m Represents the lower limit of the difference range, A n Represents the upper limit of the difference range. The upper and lower limits of the weight distribution ratio refer to the two boundary values ​​that determine the range of the weight distribution ratio. mRepresents the upper limit of weight allocation ratio, W n Represents the upper limit of the weight distribution ratio. The purpose of setting the upper and lower limits of the difference range is to limit the value range of the difference and ensure that the weight distribution ratio is adjusted within the set range. In the above formula: If W is calculated z >1, it means that the spatial characteristics of the surrounding monitoring points are not considered, and W is set z =1.

[0066] By setting a reasonable mapping range through Bayesian optimization and comparative verification, the weight distribution ratio of the time series prediction value and the interpolation prediction value is determined to more accurately characterize the spatial correlation between the target site and the neighboring sites, and the fusion prediction model is obtained. The model weight distribution is as follows:

[0067] D z =(1-W d_z )*D p +W d_z *D l

[0068] V z =(1-W v _ z )*V p +W v_z *V l

[0069] The specific operation of obtaining the prediction output is: according to the fusion prediction model optimized in step 10, the wind direction time series prediction value and the wind direction interpolation prediction value are input into the model to obtain the final wind direction prediction value; the wind speed time series prediction value and the wind speed interpolation prediction value are input into the model to obtain the final wind speed prediction value.

[0070] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A wind speed prediction method based on EMD-LSTM fusion spatial features, characterized in that: The steps include: Collect time series data and spatial data in the target monitoring area and mark the actual wind direction observation value D at the current moment l , the actual observed value of wind speed V t ; Fill in missing values ​​and outliers, and construct time series data sets and spatial direction components; Perform EMD decomposition on the wind speed time series in the time series data to extract wind speed components at different time scales; Combined with all time series features as input, the LSTM time series model is iteratively trained to output the prediction component, and the wind speed prediction value in the time series space is accumulated after denormalization. Use the trained LSTM time series model to predict the wind direction time series prediction value D at the next moment t And the wind speed time series prediction value V l ; The interpolation of the spatial direction component is calculated, and the wind speed is vector-reduced to obtain the interpolation prediction value D of the wind direction at the current moment. p and wind speed interpolation prediction value V p ; Calculate spatial independence weight: W d_a and W v_a ; Perform spatiotemporal Bayesian optimization on the spatial independence weights; obtain the optimized spatial independence weights W d_z , W d_z , establish a fusion prediction model to predict the wind speed vector in time and space; Output the final wind direction prediction value D z and the final wind speed prediction value V z .

2. The wind speed prediction method based on EMD-LSTM fusion spatial features according to claim 1 is characterized in that: The time series data includes monitoring time, wind speed, wind direction, temperature, air pressure, and humidity; the spatial data includes site number, longitude and latitude.

3. The wind speed prediction method based on EMD-LSTM fusion spatial features according to claim 1, characterized in that: The construction of the sequence data set and the spatial component includes: Time series data set: Generate corresponding seasonal features according to the monitoring time, perform one-hot encoding, and normalize; Spatial component: select the k neighboring stations with the closest Euclidean distance to the target monitoring station, and further decompose the wind speed vector data of all neighboring stations at that moment into wind speed components in the east-west and north-south directions.

4. The wind speed prediction method based on EMD-LSTM fusion spatial features according to claim 1, characterized in that: The wind speed sequence component characteristics at different time scales in the EMD decomposition are short-term fluctuation characteristics, periodic characteristics or local characteristics further divided into fine granularity.

5. The wind speed prediction method based on EMD-LSTM fusion spatial features according to claim 1, characterized in that: The loss function of the iterative training LSTM model is defined as: The wind speed loss is calculated using the mean square error loss function, and the wind direction loss is calculated using the cyclic loss function to process the periodicity of the angle. The overall loss is the weighted sum of the wind speed loss and the wind direction loss.

6. The wind speed prediction method based on EMD-LSTM fusion spatial features according to claim 1, characterized in that: The interpolation calculation for the spatial direction component is performed by using IDW inverse distance weighted calculation.

7. The wind speed prediction method based on EMD-LSTM fusion spatial features according to claim 1, characterized in that: The spatial independence weight of wind direction W d_a , wind speed spatial independence weight W v_a For: W d_a =|D t -D p |, W v_a =|V t -V p |.

8. The wind speed prediction method based on EMD-LSTM fusion spatial features according to claim 1, characterized in that: is the two spatial independence weights: W d_a and W v_a , respectively into the following mapping function W a Perform Bayesian optimization of spatiotemporal weights to obtain the optimized spatial independence weight W d_z , W d_z ; Among them, W a represents the independence weight term, which is used to adjust the weights of the time series prediction value and the interpolation prediction value to reflect the degree to which the target site is affected by the overall surrounding monitoring points; A m Represents the lower limit of the difference range, A n Represents the upper limit of the difference range; W m Represents the upper limit of weight allocation ratio, W n The upper limit of the proportion of representative weight distribution; In the above formula, if we calculate W z <1, then W is not correct z Make adjustments; otherwise, if W z When >1, it means that the spatial characteristics of the surrounding monitoring points are not considered, so W is set z =1.

9. The wind speed prediction method based on EMD-LSTM fusion spatial features according to claim 1, characterized in that: The fusion prediction model is: V z =(1-W d_z )*D p +W d_z *D1 V z =(1-W v_z )*V p +W v_z *V1 Among them, D z is the final wind direction prediction value, V z is the final wind speed prediction value.

Images