These authors have contributed equally to this work
This article was submitted to Environmental Informatics and Remote Sensing, a section of the journal Frontiers in Earth Science
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Wind speed forecasting is an important issue in Marine fisheries. Improving the accuracy of wind speed forecasting is helpful to reduce the loss of fishery economy caused by strong wind. This paper proposes a wind speed forecasting method for fishing harbor anchorage based on a novel deep convolutional neural network. By combining the actual monitoring data of the automatic weather station with the numerical weather prediction (NWP) products, the proposed method constructing a deep convolutional neural network was based wind speed forecasting model. The model includes a one-dimensional convolution module (1D-CM) and a two-dimensional convolution module (2D-CM), in which 1D-CM extracts the time series features of the meteorological data, and 2D-CM is used to mine the latent semantic information from the outputs of 1D-CM. In order to alleviate the overfitting problem of the model, the L2 regularization and the dropout strategies are adopted in the training process, which improves the generalization of the model with higher reliability for wind speed prediction. Simulation experiments were carried out, using the 2016 wind speed and related meteorological data of a sheltered anchorage in Xiangshan, Ningbo, China. The results showed that, for wind speed forecast in the next 1 h, the proposed method outperform the traditional methods in terms of prediction accuracy; the mean absolute error (MAE) and the mean absolute percentage error (MAPE) of the proposed method are 0.3945 m/s and 5.71%, respectively.
In the coastal areas of China, marine resources are abundant, and the local economic development mainly depends on fishing, marine transportation, marine oil, gas industry, etc. The rapid development of the marine economy brings prosperity to the local economy, but it also brings a number of safety issues, especially for small- and medium-sized fishing boats and fishermen who need shelter from strong winds. Statistics show that the majority of fishing boat windstorm accidents occur near ports, accounting for 68% of all fishing boat windstorm accidents, which is not only related to the delay in taking shelter from the wind but also related to the level of sheltered anchorage chosen by fishing boats. When the actual wind speed is greater than the wind resistance in the harbor anchorage, there is a risk of damage to the fishing vessel and loss of life to the crew. Besides, the large-scale integration of wind power and power grid requires accurate short-term wind speed prediction, especially for power system transmission and distribution planning, stability, reliability and safety prediction.
In the literature, many wind speed-forecasting approaches have been proposed to obtain reliable forecasts. Basic wind speed forecast methods are categorized into three classes, which consist of physical, statistical, and machine learning methods (
The statistical methods are the persistence method, Kalman filter, autoregressive moving average (ARMA) method, etc. (
With the rapid development of deep learning, the model based on deep neural network is better than that based on shallow model in feature extraction, and can improve the accuracy of wind speed prediction (
As shown in
The structure of the proposed wind speed prediction model.
In this model, the input data are composed of a variety of meteorological parameter time series. After the model input construction, the characteristic graph of time series suitable for deep convolutional neural network model processing can be generated. The first layer of the model is a one-dimensional convolutional layer, which uses 32 convolutional kernels of size
In addition to its autocorrelation, wind speed is also related to meteorological variables such as wind direction, temperature, and atmospheric pressure (
The deep convolutional neural network model proposed in this paper takes the time series feature maps as the input of the model, which is similar to the word vector representation method in natural language processing. The wind speed value at a certain time and various meteorological variables at the same time are connected in series to form a set of vectors, thus, forming a new time series unit. Then the time series unit, composed of the collected measured data and NWP data, is arranged in order of time; the measured and NWP data are intercepted, in turn, by means of a sliding window, and the intercepted data are combined into a feature map. The process of model input construction is shown in
The process of model input construction.
As shown, the measured data of the sliding window size are
The above content has given the basic structure of the proposed model and the construction method of the model input, and then the details of the model are elaborated:
The basic structure of 1D-CNN.
Suppose the time series of the input is
In the above formula,
Architecture of 2D-CNN.
Suppose the input feature map is
In the formula
In addition to the two-dimensional convolutional layer, the 2D-CNN in this paper also includes max pooling. The use of maximum pooling sampling will not only reduce the parameters of the model but also extract more obvious abstract features for the wind speed prediction task, so that it can be better utilized by the subsequent network layer. The max pooling layer is defined by
In
The last part of the 2D-CNN structure is the full connection layer, which further combines the global deep abstract features. The output results are as follows:
In
In
In order to train the proposed model and deal with the task of predicting wind speed in the next hour, a set of training feature maps was constructed from the time series of the original meteorological parameters, feature maps
In
In deep convolutional neural network, one of the biggest problems is prone to overfitting. In order to prevent this problem, regularization method can be used to strengthen the generalization ability of the model. In this paper, L2 norm is adopted to constrain the connection weight matrix of each layer, which can reduce the complexity of the model. The expression of the objective loss function minimized after L2 norm is adopted is:
In
After the objective loss function is given, the backpropagation algorithm is used to train the entire deep convolutional neural network model and selects the Adam algorithm (
In addition to regularization, the dropout strategy can also be used to further prevent the occurrence of overfitting of the model. The dropout strategy is an optimization during network training. Hidden neurons in the network are randomly discarded according to a certain probability. During the training process, discarded neurons do not participate in forward and backpropagation, but their corresponding weights are retained. On the one hand, this operation can effectively reduce the number of internal parameters in the model; on the other hand, it increases the diversity of input data in the model and, to some extent, reduces the probability of the occurrence of overfitting phenomenon. The dropout technology diagram is shown in
Dropout strategy schematic.
In this paper, the measured and NWP data of a harbor anchorage in Xiangshan, Ningbo, China, from January 1, 2016 to December 31, 2016, were used as experimental samples to verify the validity of the prediction model based on deep convolutional neural network, where the sampling interval of samples is 1 h. Due to the obvious seasonal characteristics of wind speed, five consecutive days in each season were randomly selected as the test period to verify the accuracy of this model in wind speed prediction and the experimental comparison with other traditional wind speed prediction models.
In order to comprehensively evaluate the predictive performance of the model, mean absolute error (MAE), mean absolute percentage error (MAPE), and root of the mean squared error (RMSE) were used as error evaluation indexes; the expressions of error indexes are as follows:
In the above formula:
In this paper, in addition to being randomly selected from each season for 5 days as a test period, and to the rest of the samples according to the proportion of
The super parameters of the proposed deep convolutional neural network model include the model structure, learning rate, and regular term coefficient. The model structure of this paper is the three-layer convolutional layer prediction model structure given in
The curve of the loss function during the training of the model.
It can be seen from
Next, in order to explain the rationality of the network structure of the proposed model, the control variable method is adopted to deepen the model structure of the 1D-CNN model and the proposed model step by step. The prediction effect of the different structure models was tested by deepening the 1D-CNN layer or 2D-CNN layer continuously, in which the regression prediction layer remained unchanged and remained the fully connected layer of the two-layer structure. Among them, the 1D-CNN model only uses the 1D-CNN layer as the feature extraction layer of the convolutional neural network. After extracting the temporal feature information with the 1D-CNN layer of the first layer, the proposed model uses the 2D-CNN layer of multiple layers to extract the deep abstract feature information. The prediction scatter diagram of the different models with different convolutional layer structures is shown in
Prediction scatter diagram of the different convolutional layer structures in the 1D-CNN model.
Prediction scatter diagram of the different convolutional layer structures of the model in this paper.
It can be seen from
Histogram of the prediction error of the two models under different convolutional layer structures.
In the model comparison experiment in this paper, the test set is derived from the random selection of consecutive 5 days in each season as the test period. In order to compare the prediction performance difference between the traditional wind speed prediction model and the proposed model, the persistence method, ARMA, SVR, DBN, LSTM, and 1D-CNN were selected as controls for wind speed prediction in the next hour.
As a benchmark model, the persistence method is the simplest wind speed prediction method, which takes the observed wind speed of the nearest point as the predicted value of the next point. This method is suitable for the prediction below 3∼6 h. The persistence method and the absolute error of the forecast results of the proposed method are given in
Comparison of the persistence method and ours.
Absolute error distributions of the persistence model and ours.
It can be seen from
Next, comparative experiments were conducted with other models. In the ARMA model, the wind speed series data are smoothed first; after parameter tuning, the parameter
As can be seen from
Prediction results of the different models.
Absolute errors of the different models.
Wind speed prediction results of different prediction models.
Models | MAE/(m/s) | RMSE/(m/s) | MAPE (%) | Training time (s) | Prediction time (s) |
---|---|---|---|---|---|
ARMA | 0.8846 | 1.1324 | 11.34 | 235.46 | 0.041 |
SVR | 0.9051 | 1.0647 | 12.82 | 247.23 | 0.036 |
LSTM | 0.6315 | 0.6627 | 8.76 | 1,166.34 | 0.231 |
DBN | 0.6531 | 0.6942 | 9.28 | 2,123.43 | 0.264 |
1D-CNN | 0.4424 | 0.4838 | 6.32 | 539.58 | 0.177 |
Proposed | 0.3945 | 0.4499 | 5.71 | 949.66 | 0.184 |
In
From the analysis of the training and testing time, the ARMA and SVR models have a faster speed in the training time, which were 235.46 and 247.23 s, respectively, to complete the training of the model. This is because these models have fewer training parameters compared with deep learning models. The deep learning model LSTM, DBN, 1D-CNN, and proposed model have more hierarchical structures and neural units, so they are accompanied by more training parameters. More training parameters will make the model more complicated and greatly increase the burden of model training. These models have more hierarchical structures and neural units, so they have a large number of training parameters to make the model more complicated and greatly increase the burden of model training. From the comparison of training and testing time of the different models, it can be seen that the proposed model has certain advantages over other traditional deep learning models. It only took 949.66 s to complete the training of the model, and it took 0.184 s to predict 480 samples forward.
To avoid the randomness of the prediction model and the contingency of the experimental results, next, we verified the proposed model on more datasets. In addition to the dataset used in this paper (Ningbo Xiangshan Dataset, referred to as NBdata), the first dataset we chose was the Eastern Wind Integration Dataset (EWIdata) (
We have calculated the mean and standard deviation of the above three datasets, and the results are shown in
The mean and standard deviation of three different datasets.
Dataset | Mean (m/s) | Standard deviation (m/s) |
---|---|---|
EWIdata | 4.084 | 2.773 |
WTdata | 3.082 | 2.176 |
NBdata | 6.251 | 4.395 |
Comparison of the average mean absolute percentage error (MAPE) of the different prediction models on different datasets.
Average MAPE of different prediction models on different datasets.
Dataset | ARMA (%) | SVR (%) | LSTM (%) | DBN (%) | 1D-CNN (%) | Ours (%) |
---|---|---|---|---|---|---|
EWIdata | 12.56 | 11.74 | 7.60 | 8.73 | 6.74 | 6.51 |
WTdata | 12.80 | 11.97 | 7.93 | 8.42 | 7.02 | 6.84 |
NBdata | 11.34 | 12.82 | 8.76 | 9.28 | 6.32 | 5.71 |
In this paper, deep convolutional neural network is applied to the prediction of wind speed in harbor anchorage, and a deep convolutional neural network model based on 1D-CNN and 2D-CNN is proposed. First, by using the model input construction method, the feature maps of the model input is constructed from the time series of each meteorological parameter, providing the model in this paper with the input data type with two-dimensional characteristics. Then, the backpropagation algorithm and Adam gradient descent algorithm were used to train the model. Finally, experimental verification was carried out in a test period of five consecutive days in randomly selected seasons. Based on the analysis of the experimental results, the following conclusions could be drawn: 1) In the proposed deep convolutional neural network model, the overfitting phenomenon easily occurs in the model after training, and the probability of this phenomenon can be reduced by L2 regularization and dropout strategy, so that the model after training can also have good predictive ability for unknown data. 2) Under the condition that the structure of the regression prediction layer does not change, the control variable method is adopted to add the CNN layer by layer to determine the structure of the deep convolutional neural network model. In the process of error analysis, it can be found that the deep convolutional neural network with 1 1D-CNN layer and 2 2D-CNN layers has the best prediction performance. 3) Compared with the traditional machine learning and deep learning models, the proposed model automatically extracts the time sequence feature information and deep abstract feature information from the input feature maps. It can effectively predict the wind speed of the harbor anchorage in the next hour, its prediction accuracy is higher than that of the traditional prediction model, and the model also has strong generalization performance on different datasets.
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
CH, QC, RF, and WJ contributed to the conception and design of study. CH, QC, XF, and YZ contributed to the model construction and data analysis. CH contributed to the measured data and NWP data of port anchorage collection. All authors approved the final manuscript.
This work was supported, in part, by the National Natural Science Foundation of China under Grant 41871285 and the Public Welfare Science and Technology Project of Ningbo under Grant 202002N3104.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors, and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.