Forecasting method of monthly wind power generation based on climate model and long short-term memory neural network

Rui Yin1,Dengxuan Li2*,Yifeng Wang1,Weidong Chen2

1.State Grid Hebei Electric Power Company,Shijiazhuang 050022,P.R.China

2.China Electric Power Research Institute,Nanjing 210003,P.R.China

Abstract:Predicting wind power generation over the medium and long term is helpful for dispatching departments,as it aids in constructing generation plans and electricity market transactions.This study presents a monthly wind power generation forecasting method based on a climate model and long short-term memory (LSTM) neural network.A nonlinear mapping model is established between the meteorological elements and wind power monthly utilization hours.After considering the meteorological data (as predicted for the future) and new installed capacity planning,the monthly wind power generation forecast results are output.A case study shows the effectiveness of the prediction method.

Keywords:Wind power,Monthly generation forecast,Climate model,LSTM neural network.

0 Introduction

In recent years,wind power has developed rapidly in China.The wind resources in China are characterized by concentrated distributions,large scales,and distances far away from the load center [1-2].Considering the complementary characteristics of the regional resources in China,wind power can be delivered to a load center through a large-scale and long-term power trade,so as to facilitate wind power consumption [3].The accurate prediction of regional mid-/long-term wind power generation is a key factor in implementing trans-regional and long-term power trades to address large-scale wind energy dissipation [4-5].

Compared with short-term wind farm power forecasting,there are some difficulties in middle- and long-term wind power monthly generation forecasting.For example,as China’s wind power only been in operation for a short period of time,there is very little historical data on wind power generation and wind farms.Moreover,owing to the rapid development of wind power in China,the installed capacity of new wind power can be large and irregular [6-8].The grey model is used to predict the monthly wind power generation in the power grid for northwest China,but the grey model cannot avoid the uncertainty caused by large amounts of new installed capacity [9-10].The grey model is used to realize the annual power forecast of a wind turbine in a wind farm.This method needs a large amount of long-term wind farm wind data,but may be feasible for performing annual power forecasts for individual wind farms.However,this method is difficult to employ for regional wind power forecasting,as not all wind farms provide complete data storage [11].The estimated annual wind power output of a wind farm is obtained by multiplying the average annual wind speed hours of each class of wind turbine by the total output power of the corresponding wind speed class.However,owing to incomplete consideration in the data analysis and processing,there is a certain deviation between the standard power curve of the wind turbine and tactual situation,leading to a large deviation between the final evaluation result and actual value [12-13].

This paper presents a wind power monthly generation forecasting method based on a climate model and long short-term memory (LSTM) neural network.The method can provide data support for dispatching departments,e.g.,for making long-term generation plans and long-term electricity market transactions.

1 Forecasting Methods

1.1 Climate model

Weather models are often used in short-term wind power forecasting,but are not suitable for monthly-scale power generation forecasting.Therefore,a climate model with a more accurate description of large-scale variation trends is introduced [14-15].

Climate and weather models focus on different objects.Weather models mainly focus on short-term changes in the atmosphere;thus,less consideration is given to landatmosphere interactions and other processes [16].In contrast,climate models must consider the interactions among the various spheres of Earth’s systems.The goal of weather research is to focus on instantaneous weather phenomena as much as possible;the focus of climate research is on the average state over certain time and space scales.The content of weather research concerns the precise evolution process of the weather,and the content of climate research concerns the balancing and conversion of energy [17-18].

The differences between climate and weather models are mainly reflected in the following aspects.First,the weather model emphasizes precision in time and space,and requires less integration time and stability.Thus,the dynamic framework can be solved for using a highprecision algorithm.Climate models require long-term integral stability,so the overall conservation of the dynamic framework is more important.Second,a weather model must accurately describe the evolution of an initial weather system,so the accuracy and coordination of the initial values are very important to the model.The climate model adapts to external conditions,so it is necessary to describe the external forced changes and feedback mechanisms within the climate system.Finally,in the vertical layer design of the models,the weather model has more dense layers near the boundary layer,and the requirement for the upper layer of the model is lower.Thus,the height of the top layer of the model does not need to be very high.In contrast,the climate model must ensure the balance of the energy budget,so it is necessary to consider the energy budget at the top of the model.The height of the top layer is higher,and the layers are denser [17-19].

This study selects the community earth system model(CESM),which is one of the main climate models used by the Intergovernmental Panel on Climate Change to issue assessment reports [20].The CESM represents a new generation of common earth system models,and is based on a series of National Center for Atmospheric Research (NCAR) global models from the past 30 years.Its predecessor is the community climate system model(CCSM).The development from CCSM to CESM represents a trend.The CESM is mainly used to study the past,present,and future climates of Earth [21-22].

The CESM is a fully coupled global climate system model,and can be used to simulate many elements of Earth’s climate system [23].The CESM includes modules for the atmosphere,land,ocean,sea ice,and land ice,along with a coupler.Each module has several independent modes,such as Cam,Datm,Xatm,Satm.Among them,Xatm is only used in the test.Datm is the data mode;in most cases,it is responsible for reading data,but does not participate in the calculation,as at this time the atmosphere is just a forced field.Cam participates in the operation,but also reads in the data and outputs the results,i.e.,the main action of the CESM as a climate model [24-25].

This study uses National Centers for Environmental Prediction data as the driving data of the CESM,and Community Land Model 4.0 as its land surface parameterization scheme.The data of the first 24 months of the model simulation are not stable;this data is often used as the spin-up time for the model.Accordingly,the simulation results from the first 24 months are excluded.

1.2 Long short-term memory (LSTM) neural network

The artificial intelligence model is trained using a large number of historical sample data,and a nonlinear mapping relationship is constructed between the input and output variables [26].Based on the trained model,the future wind power generation can be quantified [27].However,a traditional neural network does not consider the timing of the data when calculating the input and output [28-30].In this study,LSTM is selected as the deep neural network to consider the data timing.It contains an input layer,hidden layer,output layer,forgetting gate,input gate,and output gate.It also includes an information flow representing longterm memory,thereby forming a black box for the input and state output (namely,a cell).The specific structure is shown in Fig.1.

Fig.1 Cell structure in long short-term memory (LSTM) neural network

In the above,ht and ht-1 represent the hidden layer at time t and t-1,respectively;ct and ct-1 represent the longterm memory state of time t and t-1,respectively;and ct' is the preparatory information to be input into the long-term memory c.

Wxc represents the network weight value from the network input layer to the current module;Whc represents the network weight value from the output value of the memory module at time t-1 to the current module;and bc represents the deviation vector of the current memory module.

ft is the output signal of the forgetting gate.

s(·) represents the activation function of the neurons;Wxf represents the network weight value from the network input layer to the forgetting gate;Whf represents the network coefficient weight from the output value of the memory module at moment t-1 to the current forgetting gate;and bf represents the deviation vector of the current forgetting gate.

it is the output signal of the input gate.

Wxi represents the network weight value from the network input layer to the current input gate;Whi is the network weight value from the output value of the memory module at time t-1 to the current input gate;and bi represents the deviation vector of the current input gate.

ot is the output signal of the output gate.

Wxo represents the network weight value from the network input layer to the current module;Who represents the network weight value from the output value of the memory module at time t-1 to the current module;and bo represents the deviation vector of the current output module.

ht is the state value of the hidden layer at time t,and represents the output of the final long-term and short-term memory module.

1.3 Monthly wind power generation forecasting model

Based on the CESM climate model and historical climate data set,the historical meteorological information can be inversely calculated.The meteorological elements include the wind speed,wind direction,temperature,humidity,and air pressure.The theoretical power generation is normalized by the installed capacity to obtain the utilization hours of each historical month.This data is input into the LSTM neural network model together with the meteorological information,so as to establish a nonlinear mapping model between the meteorological elements and wind power monthly utilization hours.By combining this information with the meteorological data predicted by CESM for the future,a forecast value is obtained for the future wind power utilization hours.Finally,considering the new installed capacity planning,the monthly wind power generation forecast results are output.Fig.2 shows the specific process of the monthly wind power generation forecasting model.

Fig.2 Monthly wind power generation forecasting model

There are five main parameters that must be determined for the monthly wind power generation forecasting model based on LSTM neural network:the time steps of the input layer,dimensions of the input layer,number of hidden layers,dimensions of each hidden layer,and dimensions of the output variable.

The number of time steps in the input layer is equal to the length of the variable time series used for the monthly wind power generation prediction.The dimensions of the input layer represent the number of variables,whereas the dimensions of the output layer represent the number of variable labels (when it is multivariable).The number of hidden layers is the number of LSTM layers.With an increase in the number of hidden layers,the nonlinear fitting ability of the model will increase when the training samples are sufficient,but simultaneously,the complexity of the model and calculation and time costs for training will also increase.The dimensions of hidden layer should be determined based on the results of many experiments,and are usually set to three times the number of variable tags.The input variables selected in this model include the theoretical monthly power generation,installed capacity,wind speed,wind direction,temperature,humidity,and air pressure.

The dimensions of the output variables represent the number of output variables at each time.Only one variable(monthly wind power output) is needed,so the number of dimensions is set as one.

2 Case Study

Taking a wind farm connected to a grid in 2013 as an example,the theoretical power generation for each month from 2013 to 2018 was selected as the modeling sample,and the test sample was from January to December in 2019.The wind speed,wind direction,temperature,humidity,air pressure,and precipitation data of the wind farm from 2011 to 2019 were simulated by the CESM.As combined with the LSTM neural network model,the predicted power generation of each month in 2019 is output as shown in Fig.3;grey represents the predicted generation,and black represents the theoretical generation.

The accuracy (DE),mean absolute error (MAE),bias proportion (B),and variance proportion (V) indexes are used to evaluate the prediction effect.The MAE can evaluate the average amplitude of the error.B can reflect the degree of deviation between the mean value of the calculated value series and real value sequence;if B has a large deviation,it indicates that there is a large deviation between the predicted and real values.V can reflect the degree of deviation between the standard deviation of the calculated value sequence and the real value sequence;if the value is large,it indicates that the predicted value has evident variation from the real value.

Ei is the theoretical power generation in month i;is the predicted power generation in month i;n is the number of samples in the cycle;Cap is the installed capacity; is the mean value of the theoretical power generation data series;sE is the standard deviation of the theoretical power generation data series;and sE is the standard deviation of the predicted power generation data series.

Fig.3 Forecasted and theoretical wind power generation in 2019

The prediction error of the wind farm from January to December in 2019 was calculated.According to the calculation results for the evaluation indexes in Tab.1,the average MAE of each month is less than 15%,the DE is greater than 85%,the deviation rate is less than 3%,and the variance rate is less than 10%.Therefore,the effectiveness of the monthly wind power generation prediction method based on the climate model and LSTM neural network is verified.

Table1 Error statistics of wind farm forecast generation

Evaluating indicator Error statistics DE 86.39%Mean absolute error (MAE) 10.91%B 2.98%V 7.93%

3 Conclusion

This paper presents a monthly wind power generation forecasting method based on a climate model and LSTM neural network.The theoretical power generation is normalized by the installed capacity to obtain the utilization hours of each historical month.This data is input into the LSTM neural network model together with the meteorological information,so as to establish a nonlinear mapping model between the meteorological elements and wind power monthly utilization hours.In view of the meteorological data predicted for the future and new installed capacity planning,the monthly wind power generation forecast results are output.The case study shows the effectiveness of the prediction method.

Acknowledgments

This work was supported by National Key R &D Program of China “Study on impact assessment of ecological climate and environment on the wind farm and photovoltaic plants”(2018YFB1502800);Science and Technology Project of State Grid Hebei Electric Power Company “Research and application of medium and long-term forecasting technology for regional wind and photovoltaic resources and generation capacity” (5204BB170007);Special Fund Project of Hebei Provincial Government (19214310D).

Declaration of Competing Interest

We declare that we have no conflict of interest.

References

[1]Abdessaied N,Soeken M,Wille R et al (2013) Exact template matching using boolean satisfiability [C]//IEEE 43rd International Symposium on Multiple-clued Logic(ISMVL).Toyama:IEEE

[2]Stjepan P,Claude C,Sylvain G et al (2016) Evolutionary algorithms for boolean functions in diverse domains of cryptography [J].Evolutionary Computation,24(4):667-694

[3]Foley A M,Leahy P G,Marvuglia A et al (2012) Current methods and advances in forecasting of wind power generation [J].Renewable Energy,37(1):1-8

[4]Surhone L M,Timpledon M T (2010) World wind energy association [M].Betascript Publishing

[5]Nouredine D,Jalal F (2014) Seismic noise filtering based on generalized regression neural networks [J].Computers and Geosciences.69(04):1-9

[6]Croonen C,Dahl C M (2014) Accurate medium-term wind power forecasting in a censored classification framework [J].Energy,73(7):221-232

[7]Soman S,Zareipour H,Malij O et al (2010) A review of wind power and wind speed forecasting methods with different time horizons [C].North American Power Symposium(NAPS),IEEE,Arlington

[8]Wang X C,Guo P,Huang X B (2011) A review of wind power forecasting models [J].Energy Procedia,12:770-778

[9]Ernst B,Oakleaf B,Ahlstrom M L et aI (2007) Predicting the Wind [J].IEEE Power and Energy Magazine,5(6):75-80

[10]Nobre,Andre M,Severiano,Carlos A,Jr.Karthik,Shravan et al (2016) PV power conversion and short-term forecasting in a tropical,densely-built environment in Singapore [J].Renewable energy,94(8):496-509

[11]Cyril Voyant,Christophe Paoli,Marc Muselli et al (2013) Multihorizon solar radiation forecasting for Mediterranean locations using time series models [J].Renewable &sustainable energy reviews,28(12):44-52

[12]Antonanzas,J,Osorio N,Escobar R et al (2016) Review of photovoltaic power forecasting [J].Solar Energy,136(10):78-111

[13]Das,Utpal Kumar,Tey,Kok Soon,Seyedmahmoudian,Mehdi et al (2018) Forecasting of photovoltaic power generation and model optimization:A review [J].Renewable &sustainable energy reviews,81(1):912-928

[14]Dither Thevenard,Sophie Pelland (2013) Estimating the uncertainty in long-term photovoltaic yield predictions [J].Solar Energy,91(5):432-445

[15]Tanja Behrendt,Jan Kuehnert,Anisette Hammer et al (2013)Solar spectral irradiance derived from satellite data:A tool to improve thin film PV performance estimations [J].Solar Energy,98(12):100-110

[16]Mohamed T.Mito,Xianghong Ma,Hanan Albuflasa,Philip A Davies (2019) Reverse osmosis membrane desalination driven by wind and solar photovoltaic (PV) energy:State of the art and challenges for large-scale implementation [J].Renewable and Sustainable Energy Reviews,112

[17]Croonenbroeck C,Stadtmann G (2015) Minimizing Asymmetric Loss in Medium-term Power Forecasting [J].Renewable Energy,81:197-208

[18]Foley A M,Leahy P G,Mckeogh E J (2010) Wind power forecasting &prediction methods[C].IEEE 9th International Conference on Environment and Electrical Engineering,Prague

[19]Tian C,Dong L (2011) A Study of grey theory used in prediction of annual wind power generation [C].2011 international conference on electric information and control engineering,2011,China:Wuhan

[20]Damousis I,Alexiadis M C (2004) A fuzzy model for wind speed prediction and power generation in wind parks using spatial correlation [J].IEEE Trans on Energy Conversion,19(5):352-360

[21]Fugon L,Juban J,Kariniotakis G (2008) Data mining for wind power forecasting [C].European Wind Energy Conf,Brussels

[22]Andrew Kusiak,Zheng H Y.Song Z (2009) Short-term prediction of wind farm power:a data mining approach [J].IEEE Trans on Energy Conversion,24:125-135

[23]Li R,Wang Y (2008) Short-term wind speed forecasting for wind farm based on empirical mode decomposition [C].China

[24]Yan Z,Long K,Shi Z (2008) Wind speed predicted by wavelet analysis in input prediction control of wind turbine [C].Proc.of the 7th World Congress on Intelligent Control and Automation

[25]Lei C,Ran L (2008) Short-term wind speed forecasting model for wind farm based on wavelet decomposition [C].China

[26]Hamadi H M,Soliman S A (2005) Long-term/mid-term electric load forecasting based on term correlation and annual growth [J].Electric Power Systems Research,353-361

[27]Lorenz E,Hurka J,Heinemann D et al (2009) Irradiance Forecasting for the Power Prediction of Grid-Connected Photovoltaic Systems [J].IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,2(1):2-10

[28]Ernst B,Oakleaf B,Ahlstrom M L et aI (2007) Predicting the Wind [J].IEEE Power and Energy Magazine,5(6):75-80

[29]Nielsen T S,Mahsen H (2000) WPPT-a tool for wind power prediction [C].EWEA Special Topic Conference,Kassal

[30]Jonathan Collins,Jeremy Parkes (2009) Forecasting for utilityscale wind farms the power model challenge[R].UK:Garrad Hassan &Partners Ltd

Scan for more details

Received:5 July 2020/Accepted:27 September 2020/Published:25 December 2020

Dengxuan Li lidengxuan@epri.sgcc.com.cn

Rui Yin 1021207298@zju.edu.cn

Yifeng Wang 107920096@qq.com

Weidong Chen chenweidong@epri.sgcc.com.cn

2096-5117/© 2020 Global Energy Interconnection Development and Cooperation Organization.Production and hosting by Elsevier B.V.on behalf of KeAi Communications Co.,Ltd.This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Biographies

Rui Yin (1990-),male,doctor,engineer,research interests in flexible DC transmission technology,flexible AC transmission technology and new energy power generation.

Dengxuan Li (1991-),male,master’s degree,engineer,research interests in power meteorology and new energy power prediction.

Yifeng Wang (1987-),male,master’s degree,engineer,research interests in power grid dispatching management,new energy grid connection and operation consumption.

Weidong Chen (1989-),male,master’s degree,engineer,research interests in power meteorology and new energy power prediction.

(Editor Zhou Zhou)