Back propagation
neural networks with six different training algorithms
were obtained to be proper for predicting the fuel consumption.
The ANN with Levenberg–Marquardt training
algorithm with two hidden layers each having 10 neurons
presented better accuracy in simulation compared to
Table 4
Equations representing stepwise multiple ranges regression generated for estimating tractor fuel consumption and its performance parameter compared with ANN model.
Model Step Equation R2 Accuracy
(%)
Stepwise multiple ranges
regression
1. Drawbar power fc = 0.062Pd + 0.992 0.954 86.60
2. Drawbar power + PTo power fc = 0.052Pd + 0.01Pp + 0.621 0.963 86.74
3. Drawbar power + PTo power + throttle condition fc = 0.039Pd + 0.019Pp0.596t + 1.668 0.972 90.84
4. Drawbar power + PTo power + throttle condition + engine speed fc = 0.039Pd + 0.017Pp0.597t + 1.457 105s + 0.177 0.972 90.73
5. Drawbar power + PTo power + throttle condition + engine speed + total
tested weight
fc = 0.038Pd + 0.017Pp0.6t5.145 106s + 1.943 105wt + 1.689 0.973 91.00
ANN – 0.986 93.77
Pd: drawbar power, Pp: PTO power, t: throttle conditions and s: engine speed.
Estimated data of fuel consumption (gal h-1)
Actual data of fuel consumption (gal h-1)
Fig. 4. Actual data (From NTTL) versus estimated data of tractor fuel
consumption. Solid lines are the 1:1 relationship; dotted lines are ±CV
around the 1:1 line.
0 5 10 15 20 25
Estimated data of fuel consumption (gal h-1)
Actual data of fuel consumption (gal h-1)
Fig. 5. The diagram representing the overestimated and underestimated
data of fuel consumption.
2110 F. Rahimi-Ajdadi, Y. Abbaspour-Gilandeh / Measurement 44 (2011) 2104–2111
others