Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Solar Power Forecast for Energy Management Systemsin Smart Grid
Thailand-China Joint Public Seminar onSmart Grid and Renewable Energy Forecast
Naebboon Hoonchareon, David Banjerdpongchai
Wanchalerm Pora, Suchin Arunsawatwong, Jitkomut Songsiri
Department of Electrical Engineering
Outline
• Chula team and collaborators
• resources and facilities
• forecasting models
– nowcasting– very short-term forecasting– short-term forecasting
• methods and implementation
Collaboration
Thailand-China Research Collaboration on Renewable Energy (Year 2/3)
1
Chula team members Professors
professors: [email protected], [email protected],[email protected], [email protected], [email protected]
2
Project overview
we propose models for mainly three forecasting horizons
term horizon time step (min) applications
nowcasting 5,10,15 min 5,10,15 frequency and voltage regulation
very short-term 4 hrs 30 scheduling, load following
short-term 24 hrs 60 planning
4
Project overview
Local monitoring system
Global forecast system (GFS)
Total sky imager(TSI)
Numerical Weather
Prediction
weather and irradiance
data pre-processing
Cloud estimation
forecasted irradianceforecasted power
Neural Network
irradiance-to-powerconversion model
Neural Network
Forecasting methods
Neural network
Model Output Statistics
Time series Nowcasting5,10,15 min
ahead
Very short-term4-hr ahead
Short-term1-day ahead
5
Resources
available equipment: rooftop PV system (8+15 kw) + pyrometer
with sensors of: wind speed, wind direction, temperature, relative humidity, UVindex, irradiance, power of solar cells
6
Available data of CU location
measurements: sampling period (mostly) is 3 mins
1. solar data:
• solar irradiance• solar power
2. weather data:
• temperature• relative humidity• wind speed• wind direction• UV index
1/1/18 2/1/18 3/1/18 4/1/18 5/1/180
500
1000
I(W
/m2)
Measured solar irradiance
1/1/18 2/1/18 3/1/18 4/1/18 5/1/1840
50
60
70
RH
(%
)
Measured relative humidity
1/1/18 2/1/18 3/1/18 4/1/18 5/1/1820
30
40
50
T(°
C)
Measured temperature
1/1/18 2/1/18 3/1/18 4/1/18 5/1/180
5
UV
in
de
x
Measured UV index
1/1/18 2/1/18 3/1/18 4/1/18 5/1/180
10
20
WS
(m
/s)
Measured wind speed
1/1/18 2/1/18 3/1/18 4/1/18 5/1/180
500
1000
I (W
/m2)
Predicted solar irradiance from WRF model
prediction: running on a PC: Intel Xeon E3-1225 3.3GHz, Ram 8GB, HDD 1TB
• hourly forecasted WRF of weather and solar irradiance (on grid 3× 3 km2)
7
Forecasting methods
method: forecast solar power from solar irradiance (main input)
t+k*1hr t+10hrs
t+k*30min t+4hrs
t+5mins t+10mins t+15minstime
time
time
Nowcasting
very short-term
short-term
ANN with cloud detection
ANN
NWP and bias correctionusing regression models
8
Nowcasting Scheme
goal: provide 3 models with forecast horizon of 5,10,15 mins
forecasting time: t moves in one minute
t+5mins t+10mins t+15mins
time timetime
Cloud detection(Threshold technique, and
Homomorphic filtering)
Cloud motion estimation(Sector method)
Cloud fraction calculation (Grid cloud fraction
method)
Irradiance forecast(ANN models)
Cloud decision imageClusters of clouds moving to the sun Cloud fraction Predicted irradiance
Weather data
Irradianceto power
Predicted power
data: weather, irradiance and cloud images from UC Sandiego, USA
9
Nowcasting Cloud detection
(a) (b) (c) (d) (e)
method: homomorphic filtering toreduce brightness around the sun
row 1: clear skyrow 2-4: overcastrow 5-6: partly cloudy
(a) sky image
(b) ground truth image
(c) proposed method
(d) thresholding NBRR method
(e) thresholding RBD method
10
Nowcasting Cloud motion
use sector method to detect whichcloud section is moving toward thesun
(a) image at time t− 2
(b) image at time t− 1
(c) image at time t
(d) constructing the grid (100x150pixel) in the direction of cloudmotion
calculate the cloud fraction on the grid of cloud moving toward the sun
cloud fraction =total cloud pixels
total cloud pixels+ total sky pixels
11
Nowcasting Solar irradiance
apply ANN to forecast solar irradiance using 10 inputs:
• cloud information at grids of interest
– cloud fraction at time t– rate of change of cloud fraction at time t
• irradiance measurements: (0,5,10,15 min lags)
– I(t)– I(t− 5)– I(t− 10)– I(t− 15)
• weather data at time t
– surface/ambient temperatures
– wind speed, wind direction
– relative humidity
12
Nowcasting Solar irradiance
forecasting results are reported on the periods of
• 11:00 to 14:30
• 10:00 to 11:00 and 14:30 to 16:00
• 9:00 to 10:00 and 16:00 to 17:00
as the number of cloud grids are different in those periods
three sets of inputs for ANN are considered
• irradiance measurements only
• irradiance and cloud information
• transformed variables (from ten) by PCA
only data of partly cloudy days were trained in ANN
13
Nowcasting Solar irradiance
RMSE of irradiance forecasts (averaged over all time periods)
sets of inputs 5-min model 10-min model 15-min model
irradiance measurements only 72.51 88.27 93.40
irradiance and cloud information 72.40 87.53 92.08
transformed inputs by PCA 72.10 86.65 89.30
• gradually decreased errors in 10-min and 15-min models when using cloud
14
Nowcasting Solar irradiance
example of forecasts from models using PCA inputs
0 10 20 30 40 50 60 70 80 90 100200
400
600
800
1000
Irrad
ianc
e (W
/m2 ) 5-min ahead forecasting
0 10 20 30 40 50 60 70 80 90 100200
400
600
800
1000
Irrad
ianc
e (W
/m2 ) 10-min ahead forecasting
0 10 20 30 40 50 60 70 80 90 100Sample of validation set
200
400
600
800
1000
Irrad
ianc
e (W
/m2 ) 15-min ahead forecasting
data are validated and plotted from partly cloudy days only
15
Very short-term forcasting Scheme
goal: provide solar power forecasts of 30-min time step with horizon of 4 hours
forecasting time: t moves every 30 mins
t+k*30min t+4hrs
time
k=1,2,...,8
Trending Irradiance data
solar powerforecast
irradiancemeasurement
ANN/RNN
local measurement
error analysis
data: weather, irradiance, solar power from solar farm in Mae Hong Son
16
Very short-term forcasting Inputs for RNN
goal: select relevant inputs for recurrent neural network
• local measurement: at current date (d) and time (t)
– irradiance I(t)– temperature T(t)– solar power P (t)
• estimates of one-day-lagged solar irradiance at future times
I(t), I(t+ 30), . . . , I(t+ 240) t is in minute
RNN target: solar power at current date and future times
P (t), P (t+ 30), . . . , P (t+ 240) t is in minute
17
Very short-term forcasting Trending irradiance
goal: estimate one-day-lagged solar irradiance using past measurements
I(d− 1) = βI(d− 1) + (1− β)I(d− 2), 0 < β < 1
known as exponential moving average
0
100
200
300
400
500
600
700
1 5 9 13 17 21 25 29 33 37 41 45
Irra
dian
ce (W
/m2)
measurement
estimates
time points
the estimates are evaluated at t = 30, 60, . . . , 240 minutes (8 time points)
18
Very short-term forcasting Solar power
forecasting result with install capacity 500 kW
time of forecast 1 hour 2 hours 3 hours 4 hoursmean of absolute error (kW) 19.5 22.4 23.5 24.5max of absolute error (kW) 153.7 160.4 197.3 199.5
histogram of forecasted errors
0.000
0.050
0.100
0.150
0.200
0.250
0.300
0.350
0.400
0-10
kW
10-2
0 kW
20-3
0 kW
30-4
0 kW
40-5
0 kW
50-6
0 kW
60-7
0 kW
70-8
0 kW
80-9
0 kW
90-1
00 k
W
100-
110
kW
110-
120
kW
120-
130
kW
130-
140
kW
140-
150
kWsummer season rainy season
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4 winter season
more diversion of forecasted errors in rainy and winter (as expected)
19
Short-term forcasting Scheme
goal: provide solar power forecasts of 1-hr time step with horizon of 10 hrs
forecasting time: t moves daily
t+k*1hr t+10hrs
time
k=1,2,...,10
Local measurements
Kalman filter
Irradiance-Power conversion
Regression model
WRF forecasting
Define WRF domain
GFSmodel
Global measurements
data: WRF forecasts, weather and irradiance at Chulalongkorn university
20
Short-term forcasting Model output statistics
MOS is a multiple linear regression (regress I on relevant variables)
• model 1: 6 regressors
I(t) = β1Iwrf(t)+β2RHwrf(t)+β3Twrf(t)+β4Iclr(t)+β5cosθ(t)+β6kwrf(t)
• model 2: 3 regressors
I(t) = β1Iwrf(t) + β2RHwrf(t) + β3Twrf(t)
goal: use MOS to improve predicted I from WRF forecasts and clear sky model
(10-step prediction limits us from using local measurements as regressors)
21
Short-term forcasting Kalman filter
MOS with varying regression coefficients
I(t) = β1(t)x1(t) + β2(t)x2(t) + · · ·+ βn(t)xn(t)
regression in state-space form:
β(t+ 1) = β(t) + w(t), I(t) =[x1(t) · · · xn(t)
]β(t) + v(t)
and use KF to estimate the state (β(t)) and output (I(t))
• assume random walk model on the regression coefficients
• process and measurement noise covariances are estimated from residual errors
22
Short-term forcasting hourly/daily models
both MOS and MOS+KF are implemented in two time scales
t+k*1hr t+10hrs
time
k=1,2,...,10
time
time
time
7AM
8AM
4pm
7AM 7AM
8AM 8AM
4pm 4pm
Hourly Model
Daily Models hourly: single model and t runshourlydaily: 10 models and t runs daily
trade off between model complexityand number of available data
23
Short-term forcasting Solar irradiance
comparison between MOS and MOS+KF models using 3 regressors
Evaluated on training data
hourly model daily model100
120
140
160
180
RM
SE
(W
/m2) MOS
KF1KF2KF3
Evaluated on validation data
hourly model daily model100
120
140
160
180
RM
SE
(W
/m2) MOS
KF1KF2KF3
• performances of MOS and MOS+KF are unexpectedly very close
• no improvement of daily models on hourly model
24
Short-term forcasting Solar irradiance
example of forecasting result using hourly model during Mar 1-16, 2018
1 2 3 4 5 6 7 8
Date
0
500
1000
W/m
2
9 10 11 12 13 14 15 16
Date
0
500
1000
W/m
2
25
Summary as of May 7, 2018
• weather data: mostly available; some data require checking and cleaning
– solar irradiance and solar power measurement are not always coherent– some samples of solar irradiance suddenly drop at noon on clear days
• WRF prediction: limited and need more time to acquire more samples sinceDec 2017
• nowcasting: using cloud info as prior improves the forecasts over usingweather and irradiance measurements but the improvement is not significant
• very short-term forecasting: RNN provides less forecasted errors than ANN
• short-term forecasting:
– MOS+KF should outperform MOS when it is running over time– daily models should outperform hourly model as it can adapt to specific
characteristic of solar irradiance at specific time– need more WRF forecasts to validate the comparisons among models
26
Acknowledgement
special Thank to
for providing technical contents in this presentation
27