DEVELOPMENT OF DPS FOR AN FPSO AND A SHUTTLE TANKER BY NEURAL NETWORKS
Eiichi Kobayashi(Kobe University of Mercantile Marine, Japan)
Masami Matsuura, Katuya Daigo, and Ikuo Yamamoto(Mitsubishi Heavy Industries, Ltd., Japan)
Hiroaki Hirayama, Masaru Ihara, and Yoshifumi Suehiro(Japan National Oil Corporation, Ltd., Japan)
Abstract: Recently, ocean oil field development is tending to expand to deep sea and/or severe external force areas in search of new petroleum resources. To maintain a high operating rate for floating oil production systems, which are promising in terms of deep-sea areas, it is important to reduce the relative horizontal distance and hawser tension between the FSPO and the shuttle tanker in tandem offloading configurations even in severe sea conditions. In this study, neural network control technology has been developed, thus enabling reduction in hawser tension by controlling the yaw angle of the FPSO. The control algorithm has been confirmed to raise the wave height limit of work from 3.5m to 4.5m through computer runs and model tests in an experimental tank.
In recent years, offshore oil field development has tended to expand toward deep-sea areas and regions where weather conditions are quite severe. Floating type oil production systems are considered to be promising in such deep-sea areas.
In the floating type tandem offloading system consisting of an FPSO and a shuttle tanker, the hawser for conveying oil from the FPSO (floating production, storage and offloading system) to the shuttle tanker may become separated, or offloading work may be interrupted to avoid hawser fracture under severe weather conditions. In light of total production costs, it is important to keep the operating ratio of this offloading work sufficiently high.
The tension that acts on a hawser in the offloading from the FPSO to the shuttle tanker can be lowered even under severe weather conditions by performing relative position control between the two, and it becomes possible to raise the overall operating ratio.
Since usually shuttle tankers do not have actuators other than the propeller and rudder for the performance of dynamic positioning control with respect to coupled bodies in a tandem condition, the actuator for the entire system is a thruster positioned at the stern side of the FPSO.
The actuator of the FPSO must be controlled so that the hawser tension, determined by motion related to factors such as the positions and yaw directions of the FPSO and the shuttle tanker, is held to within the permissible limit.
Moreover, displacement of the FPSO and the shuttle tanker changes at the end stage of off-loading, as compared with the beginning. That is, the mass of each body changes.
Furthermore, since the heading angle of the FPSO is controlled in the direction of disturbance, it is assumed that the external forces such as wind, wave and current acting on the shuttle tanker, which is located on the downstream side of such forces, are sometimes affected by the position of the FPSO.
In such a complex system, it is difficult to arrive at a mathematical model expressing the motions of the FPSO and the shuttle tanker in advance, and it is thought that there are limits to the conventional controlling method that fixes the control pattern.
However, control performance can be improved using neural network control logic for systems in which mathematical expression and modeling are not easily defined.
Accordingly, in the research reported here, a neural network application is investigated, involving the technique of controlling two floating bodies by commanding a single actuator in the system (i.e., the rotable thruster of the FPSO) by means of motion data obtained from the FPSO and the shuttle tanker.
First, this paper describes the structure of the fundamental control algorithm that can be added by means of neural networks to a conventional PID control system and can compensate dynamic positioning performance.
Next, investigations were carried out through the time domain computer simulation of the dynamic positioning system of the coupled bodies using neural networks under external force conditions such as waves and current. These included issues such as deciding the learning coefficients, adding control algorithm coefficients by means of neural networks, and adjustment of PID control gains.
Finally, model ship experiments were carried out in an experimental tank under such external force conditions to evaluate and validate the dynamic positioning technology for lowering the hawser tension by means of neural networks. It is shown that the wave height limit for offloading work is increased from 3.5m to 4.5m by application of this technology.
2. CONTROL LOGIC
Considering conventional PID control compatibility and the prevention of unexpected neuro-control, the fundamental control logic was implemented in a form that compensates conventional PID control as shown in Fig.1.
Fig.1 Compensated type control system
The reliability of the control system is secured by applying the neural control system as an auxiliary, since the circuit can be separated and the control system can operate as a conventional PID control system, even when unexpected command arises in the neural control.
The fundamental neural network expression in a given time between neurons is expressed generally as follows,
where, Wij is the synaptic weighting coefficient between neurons, Ui(t) is the value of the i-th neuron in the t-th layer, and n is the number of neurons in each layer.
The output function of a neuron, f(x) , is a sigmoid function as expressed by
f(x) = tan-1(x) (2)
The input layer consists of two neurons, the transverse deviation between the FPSO and the shuttle tanker U1
and its differentiation value U2
. The hidden layer consists of three neurons in consideration of the simplicity of neural networks. The output layer consists of a neuron U1
(3) corresponding to the required thrust of the FPSO thruster.
The terms of U1(1) and U2(1) in the input layer expressed in Formula (1) also change constantly with the progress of time.
In addition, the back propagation method, which minimizes the errors between a) neural network output values for certain sets of input values and b) desired output values for the given input values, was adopted as the decision method for the weighting of coefficient Wij. The amount of compensation of the weighting coefficients can be expressed using the following formula,
ΔWij = -ηf'(Uj(t+1))Ui(t) (3)
where η denotes learning coefficient.
Finally, the composition of the neural network control logic is as shown in Fig. 2.
Fig.2 Neural network composition for control