Power saving in radio access network
To maximize power saving in a radio access network comprising cells, an optimal action amongst actions comprising switching on one or more cells, switching off one or more cells, and doing nothing is determined using a trained model, which maximizes a long term reward on tradeoff between throughput and power, the trained model taking as input a load estimate. The trained model may be updated online using measurement results on load, throughput and power consumption.
1. An apparatus comprising:
at least one processor; and
at least one memory including computer program code, the at least one memory and computer program code configured to, with the at least one processor, cause the apparatus at least to perform:
determining, for a group of cells in a radio access network, an optimal action, using a first trained model, which is based on reinforcement learning and maximizes a long term reward on tradeoff between throughput and power saving within the group of cells, the first trained model taking as input a state, wherein the optimal action is one of actions comprising at least modifying power settings of one or more cells, switching on one or more cells, switching off one or more cells, and retaining the current cell statuses in cells of the group of cells, and wherein the state comprises at least one of a load estimate and, per a cell in the group of cells, a current cell status;
causing the optimal action to be performed in response to the optimal action being modifying power settings of one or more cells, or switching on one or more cells, or switching off one or more cells, and
applying, after an optimal action that is either switching on one or more cells or switching off one or more cells is caused to be performed to one or more cells, per a cell of the one or more cells, a freeze time, wherein during the freeze time switching on the one or more cells, or switching off the one or more cells is not possible.
2. The apparatus of claim 1 , wherein the at least one memory and computer program code are configured to, with the at least one processor, cause the apparatus further at least to perform:
receiving load and performance metrics of cells that are switched on, and power consumed by the cells that are switched on; and
updating the first trained model in response to the receiving load and performance metrics of cells that are switched on, and power consumed by the cells that are switched on.
3. The apparatus of claim 1 , wherein the at least one memory and computer program code are configured to, with the at least one processor, cause the apparatus further at least to perform the determining in response to receiving, as a new load estimate, a new load prediction from a second trained model comprised in the apparatus or in another apparatus, the second trained model outputting periodically, using at least measured load data from the radio access network as input, load predictions.
4. The apparatus of claim 1 , wherein the at least one memory and computer program code are configured to, with the at least one processor, cause the apparatus further at least to perform:
instantiating and running the first trained model as a service on top of a radio intelligent controller near real time platform; and
using a data write application programming interface of the radio intelligent controller near real time platform, when causing the optimal action to be performed.
5. A method comprising:
determining, for a group of cells in a radio access network, an optimal action, using a first trained model, which is based on reinforcement learning and maximizes a long term reward on tradeoff between throughput and power saving within the group of cells, the first trained model taking as input a state, wherein the optimal action is one of actions comprising at least modifying power settings of one or more cells, switching on one or more cells, switching off one or more cells, and retaining the current cell statuses in cells of the group of cells, and wherein the state comprises at least one of a load estimate and, per a cell in the group of cells, a current cell status;
causing the optimal action to be performed in response to the optimal action being modifying power settings of one or more cells, switching on one or more cells, or switching off one or more cells; and
applying a freeze time after an optimal action that is either switching on one or more cells or switching off one or more cells is caused to be performed, wherein during the freeze time switching on the one or more cells or switching off the one or more cells is not possible.
6. The method of claim 5 , further comprising:
receiving load and performance metrics of cells that are switched on, and power consumed by the cells that are switched on; and
updating the first trained model in response to the receiving load and performance metrics of cells that are switched on, and power consumed by the cells that are switched on.
7. The method of claim 5 , the method further comprising performing the determining in response to receiving, as a new load estimate, a new load prediction from a second trained model comprised in the apparatus or in another apparatus, the second trained model outputting periodically, using at least measured load data from the radio access network as input, load predictions.
8. The method of claim 5 , the method further comprising:
instantiating and running the first trained model as a service on top of a radio intelligent controller near real time platform; and
using a data write application programming interface of the radio intelligent controller near real time platform, when causing the optimal action to be performed.
9. The method of claim 5 , the method further comprising using Q learning as the reinforcement learning.
10. A non-transitory computer readable medium comprising program instructions for causing an apparatus to perform at least one of a first process and a second process,
wherein the first process comprises at least:
determining, for a group of cells in a radio access network, an optimal action, using a first trained model, which is based on reinforcement learning and maximizes a long term reward on tradeoff between throughput and power saving within the group of cells, the first trained model taking as input a state, wherein the optimal action is one of actions comprising at least modifying power settings of one or more cells, switching on one or more cells, switching off one or more cells, and retaining the current cell statuses in cells of the group of cells, and wherein the state comprises at least one of a load estimate and, per a cell in the group of cells, a current cell status; and
causing the optimal action to be performed in response to the optimal action being modifying power settings of one or more cells, or switching on one or more cells, or switching off one or more cells,
wherein the second process comprises at least:
initializing a first trainable model, which maximizes a long term reward on tradeoff between throughput and power saving in a radio access network comprising cells and which first trainable model outputs an optimal action, wherein the optimal action is one of actions comprising at least modifying power settings of one or more cells, switching on one or more cells, switching off one or more cells, and retaining the current cell statuses;
acquiring historical data comprising a plurality of time series of evolution of at least load data, power consumption data, and cell throughput data in the radio access network, time series comprising a plurality of time steps;
training the first trainable model to a first trained model using reinforcement learning and iterating the plurality of time series and by iterating, per a time series, the plurality of time steps; and
applying a freeze time after an optimal action that is either switching on one or more cells or switching off one or more cells is caused to be performed, wherein during the freeze time switching on the one or more cells or switching off the one or more cells is not possible.
11. An apparatus comprising:
at least one processor; and
at least one memory including computer program code, the at least one memory and computer program code configured to, with the at least one processor, cause the apparatus at least to perform:
initializing a first trainable model, which maximizes a long term reward on tradeoff between throughput and power saving in a radio access network comprising cells and which first trainable model outputs an optimal action, wherein the optimal action is one of actions comprising at least modifying power settings of one or more cells, switching on one or more cells, switching off one or more cells, and retaining the current cell statuses;
acquiring historical data comprising a plurality of time series of evolution of at least load data, power consumption data, and cell throughput data in the radio access network, time series comprising a plurality of time steps;
training the first trainable model to a first trained model using reinforcement learning and iterating the plurality of time series and by iterating, per a time series, the plurality of time steps;
determining, for a group of cells in the radio access network, the optimal action, using the first trained model, which is based on reinforcement learning and maximizes the long term reward on tradeoff between throughput and power saving within the group of cells, the first trained model taking as input a state, wherein the state comprises at least one of a load estimate and, per a cell in the group of cells, a current cell status;
causing the optimal action to be performed in response to the optimal action being modifying power settings of one or more cells, or switching on one or more cells, or switching off one or more cells; and
applying a freeze time after an optimal action that is either switching on one or more cells or switching off one or more cells is caused to be performed, wherein during the freeze time switching on the one or more cells or switching off the one or more cells is not possible.
12. A method comprising:
initializing a first trainable model, which maximizes a long term reward on tradeoff between throughput and power saving in a radio access network comprising cells and which first trainable model outputs an optimal action, wherein the optimal action is one of actions comprising at least modifying power settings of one or more cells, switching on one or more cells, switching off one or more cells, and retaining the current cell statuses;
acquiring historical data comprising a plurality of time series of evolution of at least load data, power consumption data, and cell throughput data in the radio access network, time series comprising a plurality of time steps;
training the first trainable model to a first trained model using reinforcement learning and iterating the plurality of time series and by iterating, per a time series, the plurality of time steps;
determining, for a group of cells in the radio access network, the optimal action, using the first trained model, which is based on reinforcement learning and maximizes the long term reward on tradeoff between throughput and power saving within the group of cells, the first trained model taking as input a state, wherein the state comprises at least one of a load estimate and, per a cell in the group of cells, a current cell status;
causing the optimal action to be performed in response to the optimal action being modifying power settings of one or more cells, switching on one or more cells, or switching off one or more cells; and
applying a freeze time after an optimal action that is either switching on one or more cells or switching off one or more cells is caused to be performed, wherein during the freeze time switching on the one or more cells or switching off the one or more cells is not possible.