Professional Documents
Culture Documents
http://move.dei.polimi.it/
Analysis and development of a prediction algorithm for an electric vehicle range estimation
Supervisor: Sergio SAVARESI Co-supervisor: Matteo CORNO Milan 25/7/2012
Intro
Electric Vehicles (EV) Pros: Indipendent from fossil fuel (ecologic) Much cheaper re-fueling (economic) Better engine efficency Cons: Batteries have much lower energy density than fossil fuel (consequent lower range for EVs) Re-fueling (re-charging) takes much more time for batteries than fossil fuel (9 h vs ~3 min) Solutions: Predict range to plan mobility and avoid useless re-fueling Precise predictions helps exploiting 100% of battery capacity
Agenda
Runtime acquisition of vehicle signals Current Voltage Velocity State of charge estimation Residual range calculation
Index
Energetic characterization
Prediction algorithm
Electric Vehicle
Two-seater pure electric vehicle Asynchronous three phase power unit Rear wheel drive Max power: 15 kW (~20.4 CV) Max torque: 150 Nm Max speed: 100 km/h Declared range: 140 km Weight: 542 kg Lenght: 2880 mm Width: 1560 mm Height: 1425 mm
Setup
Signal acquisition from ECU Data sent via Bluetooth Data logging with iPhone Offline data analysis on PC
Fsend = 5 Hz
Index
Energetic characterization
Prediction algorithm
Wheel F.
Drag F.
Viscous Friction F.
Inertial F.
Braking F.
Rolling Friction F.
Slope F.
Wheel F.
Drag F.
Viscous Friction F.
Inertial F.
Braking F.
Rolling Friction F.
Slope F.
10
Wheel F.
Friction F.
Inertial F.
Braking F.
Slope F.
11
Wheel F.
Friction F.
Inertial F.
Braking F.
Slope F.
& = v2 v Mv
12
Nine coasting down trials done, decelerating from 20 m/s (~72 km/h)
Speed during each coasting down trial 20 18 16 14 Speed [m/s] 12 10 8 6 4 2 0 data1 data2 data3 data4 data5 data6 data7 data8 data9
10
20
30 Time [s]
40
50
60
13
[ N s2 / m2 ] [ N s / m] [ N]
14
Wheel F.
Identified Friction F.
&=0 v =0
Fb = 0
15
Wheel F.
Identified Friction F.
&=0 v =0
Fb = 0
16
Wheel F.
Identified Friction F.
&=0 v =0
Fb = 0
(v) 0 = Fw F f
Efficency estimation
=F (v) F w f
Energetic efficency
17
Constant velocity for lenghts of approximately 10 seconds Average of speed, voltage and current during the trial Different lenghts at different speeds (15-60 km/h)
Efficiency as function of speed 100 90 80 70 Efficency [%] 60 50 40 30 20 10 0 4
10 Speed [m/s]
12
14
16
Index
18
Energetic characterization
Prediction algorithm
19
Ebatt (t ) = E0 Pbatt (t ) dt
0
Drift problem solved by re-initializing the count at every vehicle startup Initialization problem solvable with a voltmetric mapping of the battery pack
Massimo Luraschi matr. 739979
20
21
80
75
Voltage [V]
70
65
60
55
2000
4000
8000
10000
12000
14000
22
Enom: 13 kWh
Index
23
Energetic characterization
Prediction algorithm
Prediction algorithm
24
Prediction steps
Average consumption identification Remaining battery capacity estimation Foreward projection of average consumption and residual range calculation Focus: calculate average consumption
Prediction algorithm
25
Preliminary analysis about consumption Identification of factors which have impact on consumption calculation Design of an algorithm based on each factor Baseline algorithm Algorithm based on moving window to calculate consumption Algorithm based on initialization value Algorithm based on trip knowledge and classification
Prediction algorithm
26
Analysis motivation
Asses each factors contribution Justify possible implementation costs with results
How?
Sensitivity analysis on each algorithm, assumption: decoupling of each factor Prediction performance evaluation
27
[ kWh km]
[ km]
28
Prediction with trip consumption 200 180 160 Predicted Distance [km] 140 120 100 80 60 40 20 0 0
20
40
60 Distance [km]
80
100
120
29
[ km]
2 km
[%]
30
31
[ km]
2 km
[%]
32
33
Instant total consumption is very little sensible to variations as the session goes on (memory effect) Solution: evaluate consumption with a moving window
[ kWh km]
Moving window is basically a filter and as such, it introduces phase displacement: the bigger the window, the more phase displacement Bigger windows have a stronger filtering effect, risking memory effect again (limit: window size = distance in a session) Smaller windows bring to noisy and nervous prediction
34
Prediction sensitivity on fraction dimension 200 180 160 Predicted Distance [km] 140 120 100 80 60 40 20 0 0 Baseline Fraction: 3 km Fraction: 15 km Fraction: 30 km
20
40
60 Distance [km]
80
100
120
35
23
21
20
19
36
23
21
20
19
> 1%
37
23
21
20
19
38
39
Baseline algorithm concentrates biggest error during session startups Cause: Consumption calculation is based on little data, susceptible of ample variations Solution: weighted average with initialization value
[ kWh / km]
40
Criticalities:
Cinit is only needed when starting a session Bigger Winit leads to a constant prediction Smaller Winit makes Cinit useless Winit and Wtrip must act as oblivion coefficients Good start: fix Wtrip = s(t), elapsed distance gets bigger so it gains more importance during session Winit = ?
Massimo Luraschi matr. 739979
41
Prediction sensitivity on cumulative consumption weight 180 Baseline weight: 10 km weight: 20 km weight: 30 km weight: 50 km weight: 70 km weight: 100 km
160
120
100
80
60
40 0
20
40
60 Distance [km]
80
100
120
42
Relative prediction error in relation with cumulative consumption weight 20 19.5 19 18.5 RelativeError [%] 18 17.5 17 16.5 16 15.5 15
10 km
20 km
30 km
50 km Weight [km]
70 km
100 km
Baseline
43
Relative prediction error in relation with cumulative consumption weight 20 19.5 19 18.5 RelativeError [%] 18 17.5 17 16.5
~ 1% 16
15.5 15
10 km
20 km
30 km
50 km Weight [km]
70 km
100 km
Baseline
44
Relative prediction error in relation with cumulative consumption weight 20 19.5 19 18.5 RelativeError [%] 18 17.5 17 16.5 16 15.5 15
10 km
20 km
30 km
50 km Weight [km]
70 km
100 km
Baseline
45
Cinit Accuracy:
A generic default value is enough? Memorizing each users consumption and using it as Cinit gives better prediction performance? (implementation costs) Trials on a standard track brought evidence of average consumption variation between 0.09 and 0.140 kWh/km depending on the pilot
46
Prediction sensitivity on fraction dimension 180 Baseline cons: 0.08 kWh/km cons: 0.09 kWh/km cons: 0.1 kWh/km cons: 0.11 kWh/km cons: 0.12 kWh/km cons: 0.13 kWh/km
160
120
100
80
60
40 0
10
20
30
40
50 60 Distance [km]
70
80
90
100
110
47
Relative prediction Error in relation with history weight 20 19.5 19 18.5 RelativeError [%] 18 17.5 17 16.5 16 15.5 15
0.08 kWh/km 0.09 kWh/km 0.1 kWh/km 0.11 kWh/km 0.12 kWh/km 0.13 kWh/km
Baseline
48
Relative prediction Error in relation with history weight 20 19.5 19 18.5 RelativeError [%] 18 17.5 17 16.5
~ 3%
16
15.5 15
0.08 kWh/km 0.09 kWh/km 0.1 kWh/km 0.11 kWh/km 0.12 kWh/km 0.13 kWh/km
Baseline
49
Relative prediction Error in relation with history weight 20 19.5 19 18.5 RelativeError [%] 18 17.5 17 16.5 16 15.5 15
0.08 kWh/km 0.09 kWh/km 0.1 kWh/km 0.11 kWh/km 0.12 kWh/km 0.13 kWh/km
Baseline
50
51
Big difference between urban, suburban and highway consumption Hypothesis: session track knowledge (e.g.: user sets satNav) Track is split in different track types (urban, suburban and highway) Average consumption in each track type is memorized a priori for each pilot (cumulative cons. updated each session) Predicted consumption for current session: weighted average of cumulative consumption and session consumption, for each track type
52
% (t ) = i {urb,extra,high} C
(t ) s %i (t ) C i
%i (t ) s
i { urb,extra, high}
Residual range
Massimo Luraschi matr. 739979
A(t ) = s(t ) +
E(t0 ) E(t ) % (t ) C
53
20
40
60 Distance [km]
80
100
120
54
20
40
60 Distance [km]
80
100
120
55
40
MRE: 23%
30
25
MRE: 13%
20
15
10
trip knowledge
baseline
Conclusions
56
Initial goal: design of an algorithm for predicting residual range Consumption data observation and hypothesis about influent factors Hypothesis validation with quantitative analysis Result: Baseline prediction improvement: from 23% to 13% MRE
iPhone Software
57
iPhone Software
Gateway connection
58
59
Drivers profile selection from drivers list To accept the choice, change screen by pushing one of the buttons at the bottom
60
Application startup
61
No Gateway connection
62
If gateway connection is missing, the application will get to this screen, except when sending an e-mail
63
During log session, the led blinks to notify Bluetooth data receival
64
During log session, the led blinks to notify Bluetooth data receival
65
Log file is saved, when the app is restored, a new session will be appended to the same file
66
When STOP button is pressed, log session is over and log can be sent by e-mail
67
Mail screen
Just need to push Send, to send mail message, or Cancel if you want to delete the message and get back to app home screen
68
69
Draft saved
70
Draft deletion
71
When choosing to delete the draft, a double confirmation is asked When NO button is pushed, log is deleted When OK button is pushed, email sending is re-opened
Graphic screen
72
Graphic screen shows useful information about residual range and drivers consumption
73
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
State of charge at startup Drivers cumulative cons. Updated state of charge Avg consumption of session in progress
74
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
500 m section avg. cons. : red if worse than cumulative, green viceversa
State of charge at startup Avg consumption of session in progress Red if worse, green viceversa
75
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
76
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
77
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
78
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
79
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
80
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
81
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
82
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
83
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
84
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
85
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress
86
The graph shows residual range calculated with drivers cumulative consumption and consumption of session in progress