Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

2020 International Symposium on Computer, Consumer and Control (IS3C)

Short-Term Solar Power Forecasts Considering


Various Weather Variables
You-Jing Zhong Yuan-kang Wu
Department of Electrical Engineering Department of Electrical Engineering
National Chung-Cheng University National Chung-Cheng University
Chiayi, Taiwan Chia-Yi, Taiwan
allenwu@ccu.edu.tw
p122108615tt@gmail.com

Abstract—Solar generation has been developed rapidly in or energy storage control. For solar power forecasting, several
recent years. The output of solar generation systems is affected useful input data of the training model include meteorological
2020 International Symposium on Computer, Consumer and Control (IS3C) | 978-1-7281-9362-5/20/$31.00 ©2020 IEEE | DOI: 10.1109/IS3C50286.2020.00117

by various uncertain factors, such as different weather variables, sky cloud observation, cloud movement indicators
variables. If a large number of solar power systems are observed by satellite cloud images, or numerical weather
connected to the grid, the stability of the power system would be prediction models (NWP) [3]. In this work, the input data
reduced. Therefore, we must pay attentions to solar power include historical solar power measurements and various
forecasting to avoid system instability. One of the important NWP meteorological data provided by the Central Weather
factors that may affect solar power generation is the weather Bureau (CWB) [4], and several solar sites are selected as our
condition, but the meteorological data have considerable forecasting objects.
uncertainty. Therefore, the main purpose of this paper is to
identify important weather variables that affect solar power II. PROPOSED METHODS FOR SOLAR POWER
forecasting. That is, the inputs used in this work to predict solar FORECASTING
power generation focuses on numerical weather prediction
(NWP) data, which includes meteorological data such as This work collected measured solar power and solar
radiation, precipitation, wind speed, and temperature. In irradiation from 182 PV sites, but some of the measured data
addition, this work also considers different time series of input are missing or incomplete. Additionally, it is found that the
data to explore the relation among data sequences. Finally, this relationship between solar irradiation and power generation in
work used various deep learning models for solar power different areas is different, especially when the distance
forecasting. between two PV sites is large. Therefore, it is not suitable to
predict the total power generation of 182 PV sites at the same
Keywords—Solar Generation, Weather Variables, Power time since the relation between solar irradiation and power
Forecasting, formatting, Numerical Weather Prediction (NWP), would be unclear, which reduces the forecasting accuracy.
Deep Learning. This work analyzed all the collected data from the 182 PV
I. INTRODUCTION sites and selected several PV sites with complete recorded
data. These PV sites are located in Pingtung county. In these
Global warming and climate change become more serious sites, the measured irradiance is directly proportional to the
in recent years. Therefore, many countries or international power generation.
communities have continuously promoted the policy of
energy transition, which aims to minimize the use of fossil The evaluation method for forecasting model is important.
fuels and increase renewable generation. Among various This work uses three metrics, as shown in (1-3), to calculate
renewable energy resources, solar power and wind power are the forecasting errors. They include mean absolute percentage
regarded as potential power generation methods. In the past, error (MAPE), root mean square error (RMSE), and
Taiwan relied on the use of imported fossil energy. In response normalized root mean square error (NRMSE). MAPE is a
to the rising awareness of environmental protection, Taiwan measure for the prediction accuracy of a forecasting method.
also proposed an energy transition policy, which includes the RMSE represents the standard deviation of the residuals
reduction of coal-fired generation, development of renewable (prediction errors), and it is a measure how spread out these
energies, and non-nuclear generation, while ensuring the residuals are. NRMSE is the statistical value calculated by
stability of power supply. The plan for energy support by 2025 normalizing the RMSE [5].
is 50% natural-gas power generation, 30% coal-fired power ଵ଴଴Ψ ௬ෝഢ ି௬೔
 ൌ σ௡௜ୀଵ ቚ ቚ (1)
generation and 20% renewable-energy power generation. ௡ ௬೘ೌೣ ି௬೘೔೙
Thus, it is estimated that the offshore wind installed capacity
σಿ
೔సభሺ௣௥௘ௗ௜௖௧೔ ି஺௖௧௨௔௟೔ ሻ

will reach up to 5.7GW or more in 2025, and the photovoltaic  ൌ ට (2)

systems have a total capacity of 20GW. ோெௌா
 ൌ (3)
௬೘ೌೣ ି௬೘೔೙
However, since solar photovoltaics are highly dependent
on the weather conditions, they are prone to great changes due
to the change of weather. Therefore, as a large amount of In this work, the NWP meteorological data were used as
renewable energy is integrated into the grid, it is easy to cause the input for solar power forecasting. There are three
voltage fluctuations and affect the stability of the regional forecasting models for NWP data, namely deterministic
power grid. Thus, accurate solar power forecasting can reduce forecasting model (WRFD), ensemble forecasting model
the impacts of solar power integration, and maintain the (WEPS) and real-time forecasting model (RWRF). WRFD is
system's supply and demand balance [2]. The main purpose of updated four times a day, and the forecast time is 120 hours
forecasting one-hour-ahead solar power generation is for with the data resolution of one hour. WEPS uses twenty
economic dispatch within one hour, rapid contingency events different models to predict the weather variables. The purpose

978-1-7281-9362-5/20/$31.00 ©2020 IEEE 432


DOI 10.1109/IS3C50286.2020.00117

Authorized licensed use limited to: East Carolina University. Downloaded on June 22,2021 at 09:41:18 UTC from IEEE Xplore. Restrictions apply.
of WEPS is to consider various possible weather scenarios . It
is updated four times per day and the forecast time is 84 hours
with the data resolution of three hours. RWRF is updated 24
times a day, the forecast time is 13 hours with the data
resolution of one hour. The NWP data include irradiation, air
pressure, temperature, humidity, wind speed, precipitation,
etc. In this work, irradiation, precipitable water vapor (mm),
temperature at 2m (K), total precipitation (mm), temperature Fig. Ĵ Structure of XGBoost training model
at 100m (K), water vapor mixing ratio at 100m, sea level
temperature (K), U-component wind speed at 10m (m/s) and III. FORECASTING RESULTS
V-component wind speed at 10m (m/s)[6] were selected into Based on the proposed forecasting method described in
the training model. Session II, this work implemented one-hour-ahead solar
power forecasting. Tables I-IV shows the forecasting results.
In this work, since the WEPS data are not complete and For instance, Tables I and II represent the training and
the RWRF model with a forecast time of 13 hours cannot be forecasting results using traditional ANN model, respectively.
used, only WRFD data were used for the forecasting work. Obviously, the forecasting accuracy with the inputs of all
The collected data are divided into training and forecasting NWP data and cumulative irradiation is the highest (see Table
data: the training data are taken from July to September 2019, I). Similarly, the forecasting performance is the best if the
and the forecasting data is taken from October and November. inputs to the training model include all NWP data and
Since the values of various variables are different, thus, a cumulative irradiation, as shown in Table II.
positive normalization is used for each variable, which is TABLE I: TRAINING ERRORS WITH DIFFERENT INPUTS [ANN MODEL]
shown as.
ŕraining Error
ௗ௔௧௔ି௠௔௫
‘”‘Ž‹œ‡ ൌ (4) num MAPE NRMSE
௠௔௫ି௠௜௡ input
ĩĦĪ ĩĦĪ
In this work, the inputs of the training model include actual 1 irradiation 5.6743 10.8941
solar power measurements at 1-hour and 24-hour before the 2
irradiation+Precipitableġ water
5.7059 10.9707
forecast time, i.e., P(t-1) and P(t-24), and the NWP data at vapor
3 irradiation+Temperature at 2m 5.1816 10.2104
the time of [t-24, t-1, t, t+1, t+24]. These inputs are used for 4 irradiation+Total precipitation 5.4057 10.5925
one-hour-ahead solar power forecasting. In addition to the 5 irradiation+Temperature at 100m 5.7008 11.0572
above variables, the cumulative irradiation and the 6
irradiation+Water vapor mixing
5.5410 10.5924
cumulative rainfall are also considered into the inputs of the ratio at 100m
7 irradiation+Sea level temperature 5.6609 10.8334
training model. The definition of cumulative irradiation or
irradiation+U-component wind
cumulative rainfall is the sum of irradiation or rainfall at time 8 speed at 10mġ + V-component 5.1329 10.2413
t, t-1, t-2, t+1, t+2, respectively. wind speed at 10m
In this work, the used training models include traditional 9 All NWP data 5.4743 10.4084
neural network and deep learning models. They consist of All NWP data+ cumulative
10 5.1120 9.9312
irradiation
ANN[7], LSTM[8], and XGBoost, and their typical 11 num10+ cumulative rainfall 5.3157 10.0957
structures are shown in Figs.1-3, respectively.
TABLE II: FORECASTING ERRORS WITH DIFFERENT INPUTS [ANN MODEL]
Forecasting Error
num MAPE NRMSE
input
(%) (%)
1 irradiation 5.5091 11.0934
2 irradiation+Precipitableġwater vapor 5.7092 11.2474
3 irradiation+Temperature at 2m 5.2168 11.2282
4 irradiation+Total precipitation 5.5963 11.5608
5 irradiation+Temperature at 100m 5.5495 11.0230
irradiation+Water vapor mixing
6 5.8002 11.8156
ratio at 100m
7 irradiation+Sea level temperature 6.0767 11.7384
irradiation+U-component wind
8 speed at 10m+ V-component wind 5.2664 11.2508
speed at 10m
9 All NWP data 5.7984 11.6619
Fig. 1: Structure of ANN training model All NWP data+ cumulative
10 5.3643 10.9630
irradiation
11 num10+ cumulative rainfall 5.5851 11.7291

Tables III and IV represent the training and forecasting


results using LSTM model, respectively. For the training data,
the inputs that include all NWP data and cumulative
irradiation can lead to a high forecasting accuracy. That is, the
more NWP variables you select, the better accuracy you can
Fig. 2 Structure of LSTM training model achieve. However, for the forecasting case, the forecasting

433

Authorized licensed use limited to: East Carolina University. Downloaded on June 22,2021 at 09:41:18 UTC from IEEE Xplore. Restrictions apply.
result with all NWP data and cumulative irradiation is not the TABLE VI: FORECASTING ERRORS WITH DIFFERENT INPUTS [XGBOOST
MODEL]
best, as shown in Table IV. It would be caused by over-
learning. Forecasting Error
TABLE III: TRAINING ERRORS WITH DIFFERENT INPUTS [LSTM MODEL] num MAPE NRMSE
input
(%) (%)
Training Error 1 irradiation 4.64 9.6918
num MAPE NRMSE 2 irradiation+Precipitableġwater vapor 4.49 9.6392
input 3 irradiation+Temperature at 2m 4.35 8.9617
(%) (%)
1 irradiation 6.0514 11.1603 4 irradiation+Total precipitation 4.56 9.7819
2 irradiation+Precipitableġwater vapor 6.1361 11.1205 5 irradiation+Temperature at 100m 4.34 8.9692
3 irradiation+Temperature at 2m 5.9235 10.9431 irradiation+Water vapor mixing
6 4.30 8.9697
4 irradiation+Total precipitation 5.9896 11.0604 ratio at 100m
5 irradiation+Temperature at 100m 5.9292 10.9457 7 irradiation+Sea level temperature 4.50 8.9971
irradiation+Water vapor mixing irradiation+U-component wind
6 6.0212 11.0811 8 speed at 10m+ V-component wind 4.71 9.7868
ratio at 100m
7 irradiation+Sea level temperature 6.0494 11.1496 speed at 10m
irradiation+U-component wind 9 All NWP data 4.28 8.7618
8 speed at 10m+ V-component wind 6.0883 11.1095 All NWP data+ cumulative
10 4.27 8.7681
speed at 10m irradiation
9 All NWP data 5.7848 10.6902 11 num10+ cumulative rainfall 4.25 8.7141
All NWP data+ cumulative
10 5.7558 10.7148
irradiation
11 num10+ cumulative rainfall 5.8579 10.8485 From the above forecasting results, one can conclude that
the NWP variables play an important role on solar power
TABLE IV: FORECASTING ERRORS WITH DIFFERENT INPUTS [LSTM MODEL]
forecasting. If more NWP variables are considered, the
forecasting accuracy could be improved. In addition, the
Forecasting Error forecasting model also affects the forecasting performance. If
num MAPE NRMSE a deep learning model, such as XGBoost is used, the
input
(%) (%)
1 irradiation 5.7503 10.7578
relationship among NWP data and solar power can be further
2 irradiation+Precipitableġwater vapor 6.2877 11.5504 explored, which would improve the forecasting performance.
3 irradiation+Temperature at 2m 6.0102 11.0592 Although solar irradiation is an important factor for solar
4 irradiation+Total precipitation 5.7254 10.7130 power generation, NWP irradiation is generally inaccuracy.
5 irradiation+Temperature at 100m 6.0790 11.1999 Many solar sites even do not measure solar irradiation.
irradiation+Water vapor mixing
6
ratio at 100m
5.9462 10.9027 Therefore, based on the forecasting results, if a variety of
7 irradiation+Sea level temperature 5.9485 11.0413 meteorological data can be considered into the input variables
irradiation+U-component wind of the forecasting model, solar power generation could be
8 speed at 10m+ V-component wind 5.8972 10.9365 predicted better.
speed at 10m
9 All NWP data 6.4655 12.2612
This work also selected different lead or lag time for the
All NWP data+ cumulative variables. In the above analyses, the inputs of the training
10 6.0407 11.7929
irradiation model include actual solar power measurements at 1-hour and
11 num10+ cumulative rainfall 6.0460 11.5432 24-hour before the forecast time, i.e., P(t-1) and P(t-24), and
the NWP data at the time of [t-24, t-1, t, t+1, t+24]. Next,
Tables V and VI represent the training and forecasting different time series for these variables are considered: a
results using traditional XGBoost model, respectively. It is shorter lead or lag time series is considered. For instance, the
clear that the forecasting accuracy with the inputs of all NWP inputs of the training model include actual solar power
data and cumulative irradiation is the highest. Similarly, the measurements at 1-hour before the forecast time, i.e., P(t-1),
forecasting performance is the best if the inputs to the training and the NWP data at the time of [ t-1, t, t+1]. Similar to the
model include all NWP data and cumulative irradiation (or above analyses, the cumulative irradiation and the cumulative
cumulative rainfall). rainfall are also considered into the inputs of the training
TABLE V: TRAINING ERRORS WITH DIFFERENT INPUTS [XGBOOST MODEL] model.
Table VII-XII shows the forecasting results. Based on the
Training Error results, if the considered lead or lag time is short, the
num MAPE NRMSE
input forecasting results are generally worse than those with longer
(%) (%)
1 irradiation 3.65 7.1358 lead or lag time, especially for using deep learning models. In
2 irradiation+Precipitableġwater vapor 3.20 6.2323 other words, considering more lead or lag time series for the
3 irradiation+Temperature at 2m 3.22 6.2941 NWP variables and historical solar power measurements, the
4 irradiation+Total precipitation 3.39 6.6331 training model has a high opportunity to learn the relationship
5 irradiation+Temperature at 100m 3.22 6.2692
irradiation+Water vapor mixing
among these data, leading to a higher forecasting accuracy. It
6 3.11 6.0166 is expected because the characteristics of solar power
ratio at 100m
7 irradiation+Sea level temperature 3.33 6.5018 generation in the previous day may be similar to that in the
irradiation+U-component wind present day.
8 speed at 10m+ V-component wind 2.88 5.5023
speed at 10m
9 All NWP data 2.40 4.6168
All NWP data+ cumulative
10 2.39 4.6118
irradiation
11 num10+ cumulative rainfall 2.39 4.5884

434

Authorized licensed use limited to: East Carolina University. Downloaded on June 22,2021 at 09:41:18 UTC from IEEE Xplore. Restrictions apply.
TABLE VII: TRAINING ERRORS WITH DIFFERENT INPUTS [ANN MODEL] In addition to the above forecasting cases, this work also
Train Error considered other lead or lag time series for these NWP
num MAPE NRMSE variables. If more data with longer lead or lag time are
input
(%) (%) considered into the training model, the forecasting results
1 irradiation 5.7369 10.8547
would depend on the training model. For example, if a deep
2 All NWP data 5.1791 10.0573
All NWP data+ cumulative learning-based training model such as XGBoost is applied, the
3 5.1720 10.3024 forecasting performance could be improved if more data are
irradiation
4 Num3+ cumulative rainfall 5.3289 10.1686 considered. However, it could not be the case if a traditional
neural network is applied since it cannot capture the
TABLE VIII: FORECASTING ERRORS WITH DIFFERENT INPUTS [ANN MODEL] connection among these data.
Test Error
num IV. CONCLUSIONS
MAPE NRMSE
input
(%) (%) This work has implemented one-hour-ahead solar power
1 irradiation 5.5352 11.1367 forecasts in an actual PV site. The contribution of this work is
2 All NWP data 6.3709 12.3734 to consider various NWP variables as the inputs of the training
All NWP data+ cumulative
3
irradiation
5.3377 11.5855 model. Additionally, different lead or lag time for these NWP
4 Num3+ cumulative rainfall 5.7444 11.4014 variables are also considered. In addition to the selection of
input data, different training models are also studied.
TABLE IX: TRAINING ERRORS WITH DIFFERENT INPUTS [LSTM MODEL] ACKNOWLEDGMENT
Train Error
num This work is financially supported by the Ministry of
MAPE NRMSE
input
(%) (%) Science and Technology (MOST) of Taiwan under Grant
1 irradiation 6.1553 11.3158 MOST 108-3116-F-194-001-. Project title: Development of
2 All NWP data 6.0053 11.0208 Renewable Power Forecasting Technique Combing
3
All NWP data+ cumulative
6.0129 11.0901 Numerical Weather Prediction and Artificial Intelligence.
irradiation
4 Num3+ cumulative rainfall 5.9436 11.0328 REFERENCES
[1] H. Wang, Z. Lei, X. Zhang, B. Zhou and J. J. E. C. Peng, "A review of
TABLE X: FORECASTING ERRORS WITH DIFFERENT INPUTS [LSTM MODEL] deep learning for renewable energy forecasting," Energy Conversion
and Management, vol. 198, p. 111799, 2019.
Test Error [2] J. Antonanzas, N. Osorio, R. Escobar, R. Urraca, F. J. Martinez-de-
num MAPE NRMSE Pison, and F. J. S. E. Antonanzas-Torres, "Review of photovoltaic
input
(%) (%) power forecasting," Solar Energy, vol. 136, pp. 78-111, 2016.
1 irradiation 5.7417 10.8043 [3] S. Sobri, S. Koohi-Kamali and N. A. J. E. C. Rahim, "Solar
2 All NWP data 6.0112 11.2165 photovoltaic generation forecasting methods: A review," Energy
All NWP data+ cumulative Conversion and Management, vol. 156, pp. 459-497, 2018.
3 5.9089 11.1299
irradiation [4] D. Yang, J. Kleissl, C. A. Gueymard, H. T. Pedro, and C. F. J. S. E.
4 Num3+ cumulative rainfall 5.8613 11.1462 Coimbra, "History and trends in solar irradiance and PV power
forecasting: A preliminary assessment and review using text mining,"
Solar Energy, vol. 168, pp. 60-101, 2018.
TABLE XI: TRAINING ERRORS WITH DIFFERENT INPUTS [XGBOOST MODEL]
[5] U. K. Das et al., "Forecasting of photovoltaic power generation and
Train Error
model optimization: A review," Renewable and Sustainable Energy
num MAPE NRMSE Reviews, vol. 81, pp. 912-928, 2018.
input
(%) (%)
[6] W. Lee, K. Kim, J. Park, J. Kim, and Y. J. I. A. Kim, "Forecasting solar
1 irradiation 4.26 8.3775
power using long-short term memory and convolutional neural
2 All NWP data 2.75 5.2949 networks," IEEE Access, vol. 6, pp. 73068-73080, 2018.
All NWP data+ cumulative
3 2.74 5.2991 [7] M. Q. Raza, M. Nadarajah, and C. J. S. E. Ekanayake, "On recent
irradiation
4 Num3+ cumulative rainfall 2.75 5.29844 advances in PV output power forecast," Solar Energy, vol. 136, pp.
125-144, 2016.
[8] J. Ospina, A. Newaz, and M. O. J. I. R. P. G. Faruque, "Forecasting of
TABLE XII: FORECASTING ERRORS WITH DIFFERENT INPUTS [XGBOOST PV plant output using hybrid wavelet-based LSTM-DNN structure
MODEL] model," IET Renewable Power Generation, vol. 13, no. 7, pp. 1087-
1095, 2019
Test Error
num MAPE NRMSE
input
(%) (%)
1 irradiation 4.64 9.6470
2 All NWP data 4.35 8.9276
All NWP data+ cumulative
3 4.37 8.9205
irradiation
4 Num3+ cumulative rainfall 4.35 8.8999

435

Authorized licensed use limited to: East Carolina University. Downloaded on June 22,2021 at 09:41:18 UTC from IEEE Xplore. Restrictions apply.

You might also like