Prediction of mechanical availability in forest harvesting through artificial neural networks
Name: LEONARDO CASSANI LACERDA
Publication date: 28/09/2022
Advisor:
Name |
Role |
|---|---|
| NILTON CESAR FIEDLER | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| FLAVIO CIPRIANO DE ASSIS DO CARMO | External Examiner * |
| NILTON CESAR FIEDLER | Advisor * |
| SÁULO BOLDRINI GONÇALVES | External Examiner * |
Summary: The forest sector represents a relevant part of the Brazilian economy, which is dominated for the most part by the production of plantations of the Eucalyptus sp. In general, the mechanization of activities, especially forest harvesting, is considered one of the most important in the process, adding values that sometimes exceed 50% of the total costs of the activities. Thus, the present research aimed to predict the mechanical availability of machines that constitute the harvesting systems of full-tree and Cut-to-lenght, through artificial neural networks and linear regression. The research was developed through a database from a forestry company, concentrating data such as the activities of the machines of the full-tree system in the north of the state of Espírito Santo, and the Cut-to-Length system in the south of the state of Bahia. With variables harvested per hour worked, hours for mechanics, average individual volume, trees harvested per hour, cubic meters per hour and hydraulic oil per hour, presented the mechanical prediction by means of individual mean volume artificial neural networks, having a variable in isolation with as individual artificial neural networks. Having mechanical availability as the output variable, the data were randomly divided to be used in the ANN prediction, with samples of 70% and 80% for training and 30% and 20% for validation, respectively. They were trained by three algorithms (resillient propagation, backpropagation and quick propagation) using configurations ranging from 5 to 11 neurons in the hidden layer, logistic and sigmoidal functions with 50 networks per configuration, totaling 8,400 trained networks. The linear regression analysis used only the variables that showed a significant linear correlation with productivity, according to Pearson's correlation coefficient matrix, using the t test at 5% and 1% significance. Both modeling techniques were evaluated through statistics and graphical analysis of residuals. The selected artificial neural networks presented R² above 0.95, indicating strong correlation and high accuracy in relation to the observed values. Among the training algorithms, the resilient propagation proved to be more effective in predicting the mechanical availability for both harvesting systems. In a another way, the logistic activation function predominated for the full-tree and the sigmoidal function for the cut-to-length. The average individual volume input variable, tested separately using the best configuration found in each system, showed to influence the prediction of mechanical availability. Finally, it was concluded that the prediction methodologies for Mechanical disponible were effective, with the full-tree system surpassing the Cut-to-length and for both ANNs superior to linear regression modeling.
Keywords: Artificial neural networks; Full-tree; Cut-to-length; Forest mechanization.
