Pump systems and methods to improve pump load predictions
AssigneeLandmark Graphics Corporation
InventorsSrinath Madasu, Keshava Prasad Rangarajan
Life of this patent
Estimating the term…
Abstract
The disclosed embodiments include pump systems and methods to improve pump load predictions of pumps. The method includes determining, in a neural network, a pump load of a wellbore pump based on a physics based model of the pump load of the wellbore pump. The method also includes obtaining a measured pump load of the wellbore pump. After initiation of a pump cycle of the wellbore pump, the method further includes predicting a pump load of the wellbore pump based on the physics based model, performing a Bayesian Optimization to reduce a difference between a predicted pump load and the measured pump load to less than a threshold value, and improving a prediction of the pump load based on the Bayesian Optimization.
Claims18
A method to improve prediction of a pump load of a wellbore pump, the method comprising:
- determining, in a neural network, a pump load of a wellbore pump based on a physics based model of the pump load of the wellbore pump;
- wherein the wellbore pump comprises a sucker rod and a cycle time of a pump cycle;
- obtaining a measured pump load of the wellbore pump;
- and
- after initiation of the pump cycle of the wellbore pump:
- predicting the pump load of the wellbore pump based on the physics based model, wherein the predicted pump load is λpredicted=−E*A*(y (L, t)+(u (t)−w (t)))*x, wherein E is equal to Young's Modulus, A is a cross section of a rod of the wellbore pump, y is a transformed displacement, t is time, L is a length of the rod, x is a location of a stroke of the rod, w is a rod displacement, and u is a pump displacement;
- performing a Bayesian Optimization to match a predicted pump load to the measured pump load by adjusting the cycle time of the pump cycle of the wellbore pump based on a historical value of a cycle time;
- determining a cost function of the physics based model;
- and
- improving a prediction of the pump load based on the Bayesian Optimization, wherein improving the prediction comprises applying the cost function of Cost=∫0t (λpredicted−λdetermined))2 dt to reduce a difference between the predicted pump load and the measured pump load, wherein λdetermined is the measured pump load.
A pump system, comprising:
- a wellbore pump having a sucker rod and a cycle time of a pump cycle;
- a storage medium;
- and
- one or more processors operable to:
- determine, in a neural network, a pump load of the wellbore pump based on a physics based model of the pump load of the wellbore pump;
- obtain a measured pump load of the wellbore pump;
- and
- after the pump cycle of the wellbore pump has initiated:
- predict the pump load of the wellbore pump based on the physics based model;
- perform a Bayesian Optimization to match a predicted pump load to the measured pump load by adjusting the cycle time of the pump cycle of the wellbore pump based on a historical value of a cycle time, wherein the predicted pump load is λpredicted=−E*A*(y (L,t)+(u(t)−w (t)))*x, wherein E is equal to Young's Modulus, A is a cross section of a rod of the wellbore pump, y is a transformed displacement, t is time, L is a length of the rod, x is a location of a stroke of the rod, w is a rod displacement, and u is a pump displacement;
- determine a cost function of the physics based model;
- and
- improve a prediction of the pump load based on the Bayesian Optimization,
- wherein improving the prediction comprises applying the cost function of Cost=∫0t(λpredicted−λdetermined)2 dt to reduce a difference between the predicted pump load and the measured pump load, wherein λdetermined is the measured pump load, and
- wherein operations performed to solve the physics based model, predict the pump load, perform the Bayesian Optimization, and improve the prediction of the pump load are performed in a single framework of the neural network.
A machine-readable medium comprising instructions stored therein, which when executed by one or more processors, causes the one or more processors to perform operations comprising:
- determining, in a neural network, a pump load of a wellbore pump based on a physics based model of the pump load of the wellbore pump;
- wherein the wellbore pump comprises a sucker rod and a cycle time of a pump cycle;
- obtaining a measured pump load of the wellbore pump;
- and
- after initiation of the pump cycle of the wellbore pump:
- predicting the pump load of the wellbore pump based on the physics based model, wherein the predicted pump load is λpredicted=−E*A*(y (L, t)+(u (t)−w (t)))*x, wherein E is equal to Young's Modulus, A is a cross section of a rod of the wellbore pump, y is a transformed displacement, t is time, L is a length of the rod, x is a location of a stroke of the rod, w is a rod displacement, and u is a pump displacement;
- performing a Bayesian Optimization to match a predicted pump load to the measured pump load by adjusting the cycle time of the pump cycle of the wellbore pump based on a historical value of the cycle time;
- determining a cost function of the physics based model;
- and
- improving a prediction of the pump load based on the Bayesian Optimization,
- wherein improving the prediction comprises applying the cost function of
- Cost
- =
- ∫
- 0
- t
- (
- λ
- predicted
- -
- λ
- determined
- )
- )
- 2
- dt to reduce a difference between the predicted pump load and the measured pump load, wherein λdetermined is the measured pump load, and
- wherein operations performed to solve the physics based model, predict the pump load, perform the Bayesian Optimization, and improve the prediction of the pump load are performed in a single framework of the neural network.
Dependencies are inferred from the claim text.