Model of Offshore Wind Power Forecast Considering the Variation of Wind Speed in Second-level Time Scale

To enable power generation companies to take full advantage of effective wind energy and grid companies to correctly schedule wind power, a forecast model of offshore wind power considering the variation of wind speed in second-level time scale is proposed in this paper. First, data preprocessing is utilized to process the abnormal data and complete the normalization of speed and power data. Then, a wind speed prediction model is established in the second time scale through the differential smoothing power sequence. Finally, a rolling LSTM memory network is authorized to realize the prediction of second-level time scale wind speed and power. A case of offshore wind farm is utilized to illustrate and characterize the performance of the wind power forecast model.

Abstract-To enable power generation companies to take full advantage of effective wind energy and grid companies to correctly schedule wind power, a forecast model of offshore wind power considering the variation of wind speed in second-level time scale is proposed in this paper. First, data preprocessing is utilized to process the abnormal data and complete the normalization of speed and power data. Then, a wind speed prediction model is established in the second time scale through the differential smoothing power sequence. Finally, a rolling LSTM memory network is authorized to realize the prediction of second-level time scale wind speed and power. A case of offshore wind farm is utilized to illustrate and characterize the performance of the wind power forecast model.

I. INTRODUCTION
Offshore wind power over the past decade remains unprecedented [1][2]. And offshore wind power has a pivotal and mature role in renewable energy.
High offshore wind speed, large turbine capacity, annual operating hours of up to 4,000 hours or more, offshore wind power efficiency than onshore wind power annual power generation of 20% to 40% more, with higher energy efficiency [3][4][5]. Offshore wind farms are far away from the land, not affected by urban planning, and reducing the impact of noise and electromagnetic waves on residents. Furthermore, offshore wind farms can promote the economic development of coastal areas and facilitate the local consumption of coastal heavy-load cities [6][7][8]. However, the randomness and uncontrollability of offshore wind power causing generation companies to decline to report power generation correctly which leads to active abandonment of the reported electric power and fines for excessive reported electric power [9][10]. Therefore, it is necessary to predict wind speed and predict short-term wind power in practical applications [11]. Offshore wind power prediction uses historical output data, Numeric Weather Prediction (NWP), and measured meteorological data, and a forecast model is established to predict future offshore wind power [12][13][14]. Reference [15] arrested the mining and analysis of the inherent fluctuation law of wind power as the starting point and studied new methods around the utilization of wind power fluctuation law in ultra-short-term forecasting. Reference [16] proposed a new method based on the extreme learning machine and the bootstrap prediction interval formula to predict wind power in different seasons and verify its effectiveness. Reference [17] effectively solved the combined forecasting interval of wind power by improving the bat algorithm based on the fuzzy cost function. Reference [18] employed support vector machine (SVM) regression to the model of power forecasting and effectively verified that for different wind power weather types, the neighboring days were selected to establish the reliability based on its reliability. However, wind power forecasting is urgently needed to enable power generation companies to use wind power forecasting results to plan and schedule offshore wind units and maximize the profits of power generation companies and the power sector, and it is important to accurately forecast wind power [19].
The existing multi-factor offshore wind power forecasting methods cannot satisfy the lack of information. The wind power forecast method is eagerly demanded in the next few hours when only offshore wind turbine wind speed and wind power are available. So this paper proposes a forecasting model of offshore wind power, which the variation of second-level time scale wind speed is considered. In the case of processing abnormal data and data normalization and differential smoothing power series, the LSTM rolling model is established for training, and the training results are utilized to predict the data. The training and prediction results using the real data from an offshore wind farm show that the proposed model has higher prediction accuracy compared with other traditional models. It can more accurately predict the second-level wind power in the next four hours.

A. Processing of Abnormal Data
The measurement data of offshore wind farms is collected by various sensors, and realize rapid conversion and transmission

Model of Offshore Wind Power Forecast Considering the Variation of Wind Speed in
Second-level Time Scale of field data through data transmission devices. Due to the special climatic conditions, instrument failure, network transmission errors, and other problems, may lead to wind measurement data and wind power in the collection, conversion, transmission and other processes occur in the absence of measurement and measurement error, which undermines the integrity and rationality of offshore wind power data, cannot truly reflect the wind resources distribution of wind farm. Most of the existing methods directly or indirectly use wind resource data, so the accuracy of wind measurement data is directly related to the good or bad wind power prediction results. Incomplete data reduce the coherence and utilization of the basic data, and they and the wrong data together constitute anomalous data, which affect the analysis of wind power fluctuation characteristics, the construction of prediction models and the study of prediction errors. Therefore, the original wind measurement data need to be analyzed, examined and corrected to obtain more accurate historical data.

B. Normalization of Wind Speed And Wind Power
Before training the prediction model, because the GRU neural unit in the model uses Sigmoid and tanh functions as activation functions, and also to improve the prediction accuracy of wind power and data convergence speed in the training process, this method uses the Max-Min normalization method to normalize the original wind power data and convert it to the data in the interval [0,1]. The equation of data normalization is as follows: where y is the normalized wind power value; xmax is the maximum value of raw wind power; xmin is the minimum value of raw wind power; xi is the actual wind power value. In general, the wind farm output is considered to be the superposition of each unit. The power output of wind turbine is expressed by the following equation: where Cp is the wind energy using coefficient of the unit; ρ is the air density; r is the radius of the fan blade; v is the wind speed. The wind energy using coefficient indicates the ratio of power to wind energy, which is conversion efficiency of wind energy by turbine. According to the Baez limit, the maximum wind energy using coefficient of the horizontal unit is 0.593 under the condition that not considering the influence of wake flow.

C. Fitting Relationship of Wind Speed And Wind Power
Due to the large variety of wind speed, small and excessive wind speed is not conducive to generation. Small wind speed can not drive the blade rotation, excessive wind speed will cause offshore wind turbine failure, in the design, offshore wind turbines need to install speed limiting devices to ensure that the wind turbine in high winds can operate safely. Wind turbine design has the following provisions: cut-in wind speed vin, cutout wind speed vout, rated wind speed vr. Therefore, the offshore wind power formula can also be expressed as follows: where f(v) is used as the equation of offshore wind power versus wind speed at wind speeds between the cut-in wind speed and the rated wind speed.
Even though the relationship between wind speed and wind power cannot be solved accurately in reality, and the actual equation of wind speed and wind power cannot be solved, so this paper will adopt the Sigmoidal model using the Boltzmann equation for nonlinear fitting of the actual equation of wind speed and wind power, and the Boltzmann equation is shown as follows: where A1, A2, x0, B are the parameters of the Boltzmann equation.

A. Wind Power Prediction Model in Second-level Time Scale
The time series has certain dynamic time characteristics, that is the series value of the current moment has a correlation with the series value of several previous moments, and the correlation increases with decreasing time interval. And offshore wind power has numerous uncertainties, such as wind direction, air pressure, temperature, etc., but the second-level wind speed fluctuation changes and wind power size will not occur in a large difference. According to its offshore wind fluctuation pattern and time series, the future short-term wind speed variation is judged, so as to predict the wind power size. A single wind speed variation prediction model on a secondscale time scale can be expressed as follows: where θ is the time interval of data collection; f1 is the time correlation function of the offshore wind power series; E(t) is the prediction error at moment t.
Due to the complexity of the weather system, the offshore wind power series have an unstable nature. By differentially smoothing the power series, the f1 complexity can be reduced and the prediction error can be reduced as follows: t and t-θ; f2 is the time-dependent function of the offshore wind power difference series; e(t) is the minimum prediction error at moment t. B. Rolling LSTM Neural Network Recurrent neural networks are a type of artificial neural network. Recurrent neural networks are good at processing time scale data and can describe the data before and after relationship on the time axis. LSTM was proposed by Hochreiter and Schmidhuber as a derivation of recurrent neural networks [20]. LSTM adds multiple special computational nodes in the hidden layer of recurrent neural networks to improve the gradient transfer mode during backpropagation and effectively slow down the gradient disappearance or gradient explosion, solving the problem of not being able to build a prediction model for a longer time span due to the long-term dependence problem of RNN, whose model structure is shown in Figure 2.

Fig. 2. Structure of recurrent neural network
In the figure2: U, V and W are the weight coefficient matrix; x, y and h denote the input, output and hidden layer sequences of the RNN model, respectively.
The LSTM network structure consists of input gates, output gates and forgetting gates, and is different from the RNN in that there are multiple hidden layers, and the neurons in the hidden layers are replaced by memory units with gating mechanisms. The basic structure of the network is shown in Figure 3. The memory cell is the core of the LSTM network. The input of the model contains the sequence input xt at time t, the hidden layer cell state ht-1 and the memory cell ct-1 at time t1; the output contains the memory cell state ct and the hidden layer state ht, where ct and ht each contain the long-term and short-term memory information of the model, and the information flow between the networks is carried out by controlling the input gate, the forget gates and output gates to achieve the reading and modification of the memory cell unit for information flow between networks. tanh denotes the activation function of tanh, and the input gate uses the sigmoid activation function to enter the parameters and control the variables between [0, 1] to achieve the control of xt on ct; the forgetting gate is to selectively forget the neuron state of the previous moment, and the specific expression is to use memory unit ct-1 for the control of ct; the output gate is used to output and control the parameter variables, i.e., the degree of influence of ct on ht is utilized. The calculation equations are respectively as follows: where it, ft, and ot denote the state calculation results of input gate, forget gate and output gate, respectively; Wih, Wfh, Wox and bi, bf, and bo denote the weight matrix and bias term of the corresponding gates, respectively;  represent the sigmoid activation function.
The output result of the memory module in the LSTM model at moment t is determined by the output gate together with the cell state as follows: where ct' denotes the cell state-input at moment t; tanh is the hyperbolic tangent activation function; Wc, bc represent the state weight matrix and bias term of the input layer, respectively; denotes the elements are multiplied by position.
The prediction of time series data is achieved by rolling LSTM memory network, and the specific process is shown in Figure 4.

C. Evaluation Criteria of Model
In order to accurately validate the prediction performance of proposed method based on LSTM network, the mean absolute percentage error (yMAPE) root mean square error (yRMSE) and prediction accuracy (yFA) are selected as evaluation indicators to analyze the prediction effect of the model, where the smaller values of yMAPE and yRMSE indicate a better fit and more accurate model prediction results, as defined by the formula as follows: where n denotes the sample size of the test set; Xact(i) and Xpred(i)

A. Experimental Data Set And Experimental Environment
This paper uses an experimental environment with Windows 10 operating system, 8 GB of RAM, Intel(R) Core(TM) i3-9100F CPU @ 3.60 GHz, NVIDIA GeForce GTX 1650 GPU, and Python 3.8 language. Anacaoda3 and Tensorflow1.14.0 were used to write LSTM memory networks, recurrent neural networks and ARIMA, which is commonly used for time series prediction.
To verify the scientificity and reliability of proposed method of offshore wind power prediction, the online monitoring data of second-level wind speed and second-level offshore wind power in Jiangsu province in one day are used in the analysis of this paper. 105 offshore wind units whose rated power is 1500 kW are included in this wind turbine group, and the wind speed and offshore wind power of a unit are shown in Figure 5: As can be seen from Figure 5, part of the data shows that the power is less than 0, abnormal data need to be processed, and the offshore wind power of this unit has not reached half of the rated power, so it is judged that this wind speed and power relationship should be in the rising part, and the non-linear fitting of the second wind speed and offshore wind power by Boltzmann equation is shown in Figure 6.
The difference method is used for wind power data, only the rate of change in the series is considered to exclude the trend problem that the series has, and a rolling LSTM model is used for time series prediction of offshore wind power.
The LSTM has one hidden laye. The Adam algorithm is used to train the internal parameters of the LSTM, the tanh function is used for the activation function in the hidden layer, the rounding rate of the network nodes is taken as 0.2, the number of iterations is taken as 300 to prevent overfitting, the learning rate in the LSTM model is set to 0.001, the number of neurons in the hidden layer is 4, and the first 18 hours of the day is also used for training. The last 4 hours were used for testing.

B. Results of Offshore Wind Forecast
In this paper, rolling LSTM model is selected to realize the prediction of offshore wind power in seconds, and the actual power curve and other models predict the power curve within 4 hours per second values as shown in Figure 7, and the evaluation index of the prediction results are shown in Table 1: By intercepting the data from 22:59:30 to 23:00:30, it can be seen that the LSTM second-level prediction model is closer to the real data, and the second-level power error is lower. The actual power curve under the intercepted partial time and the predicted power curve of other models are shown in Figure 8  The analysis shows that the LSTM prediction model has the lowest yMAPE, while the lowest yRMSE index and the highest yFA, respectively, compared to the RNN and ARIMA prediction models, indicating that the LSTM prediction model has better prediction results for the offshore wind power prediction problem with second-level time series wind speed variation.

V. CONCLUSION
In this paper, an offshore wind power forecast model considering the variation of wind speed in the second time series in conjunction with the current requirements of the rapid development of artificial intelligence and gradual enhancement of the offshore wind power prediction accuracy is proposed. This paper obtains conclusions as follows: (1) The LSTM rolling prediction model is used to analyze the wind speed and wind power at the second level, and the prediction of offshore wind power per second in the next 4 hours is completed.
(2) Using the characteristics of LSTM network applicable to time series, the LSTM rolling prediction model has a large enhancement in forecasting accuracy compared with RNN and ARIMA prediction models.
(3) With the rapid development of computer technology combined with the comprehensive use of big data platform, the model is applied to other prediction fields, which may uncover more effective information and thus improve the prediction accuracy, and may provide theoretical guidance for the subsequent long-term offshore wind power accurate prediction.