Dynamically forecasting high resolution air temperature in real-time using multiple sources
Techniques for dynamically forecasting high spatial resolution air temperature in real-time using a multiplicity of data sources and advanced machine learning models are provided. In one aspect, a system for generating air temperature forecasts includes: a spatial downscaling module for providing air temperature data from a multiplicity of sources as input to a generative adversarial network, conditioning the input to the generative adversarial network on at least one conditioning variable, and generating high resolution air temperature maps; and a temporal forecasting module for providing the high resolution air temperature maps along with real-time air temperature data from a sensor network as input to a physics-informed neural network with extended Kalman filter, and generating high temporal and spatial resolution air temperature forecasts. A method for generating air temperature forecasts using the present system is also provided.
1 . A computer system comprising:
a processor set;
one or more computer readable storage media; and
program instructions stored on the one or more computer readable storage media to cause the processor set to perform operations comprising:
providing air temperature data from a multiplicity of sources as a first input to a generative adversarial network;
conditioning the first input to the generative adversarial network on at least one conditioning variable;
training the generative adversarial network with the air temperature data and the conditioning the first input to the generative adversarial network on the at least one conditioning variable;
generating high resolution air temperature maps as a first output from the generative adversarial network, wherein the high resolution air temperature maps are generated at a higher spatial resolution than the air temperature data from the multiplicity of sources;
providing the high resolution air temperature maps along with real-time air temperature data from a sensor network as a second input to a physics-informed neural network with extended Kalman filter; and
generating high temporal and spatial resolution air temperature forecasts as a second output from the physics-informed neural network with the extended Kalman filter.
2 . The computer system of claim 1 , wherein the high resolution air temperature maps have a spatial resolution of 100 meters×100 meters or finer.
3 . The computer system of claim 1 , wherein the multiplicity of sources is selected from a group consisting of: global climate model forecasts, regional climate model forecasts, and urban climate model datasets.
4 . The computer system of claim 1 , wherein the first input to the generative adversarial network comprises a first timeseries of the air temperature data from the multiplicity of sources, and wherein the first output from the generative adversarial network comprises a second timeseries of the high resolution air temperature maps.
5 . The computer system of claim 1 , wherein the at least one conditioning variable is obtained from a dataset selected from a group consisting of: a satellite imagery dataset, a land cover dataset, and an elevation dataset.
6 . The computer system of claim 1 , wherein the operations further comprise:
generating the high resolution air temperature maps; and
feeding the high resolution air temperature maps to a discriminator sub-model along with urban climate model datasets corrected using the real-time air temperature data from the sensor network.
7 . The computer system of claim 6 , wherein the urban climate model datasets have a spatial resolution of 100 meters×100 meters or coarser.
8 . The computer system of claim 1 , wherein the high temporal and spatial resolution air temperature forecasts are generated at a real-time (t) or a near real-time, wherein the near real-time comprises a delay in the real-time (t) of less than about 15 minutes.
9 . The computer system of claim 1 , wherein the high temporal and spatial resolution air temperature forecasts are generated at a spatial resolution of 100 meters×100 meters or finer.
10 . A computer system comprising:
a processor set;
one or more computer readable storage media; and
program instructions stored on the one or more computer readable storage media to cause the processor set to perform operations comprising:
providing air temperature data from a multiplicity of sources as first input to a generative adversarial network;
conditioning the first input to the generative adversarial network on at least one conditioning variable;
training the generative adversarial network with the air temperature data and the conditioning the first input to the generative adversarial network on the at least one conditioning variable;
generating high resolution air temperature maps having a spatial resolution of 100 meters×100 meters or finer as a first output from the generative adversarial network, wherein the high resolution air temperature maps are generated at a higher spatial resolution than the air temperature data from the multiplicity of sources;
providing the high resolution air temperature maps along with real-time air temperature data from a sensor network as a second input to a physics-informed neural network with extended Kalman filter; and
generating high temporal and spatial resolution air temperature forecasts as a second output from the physics-informed neural network with the extended Kalman filter, wherein a loss function of the physics-informed neural network computes loss L as L=L b +L o +L e +L i , wherein L e is a residual loss, L b is a boundary condition loss, L o is an initial condition loss, and L i is a loss term defined by an innovation vector and Kalman gain in the extended Kalman filter.
11 . The computer system of claim 10 , wherein the multiplicity of sources is selected from a first group consisting of: global climate model forecasts, regional climate model forecasts, urban climate model datasets, and combinations thereof, and wherein the at least one conditioning variable is obtained from a dataset selected from a second group consisting of: a satellite imagery dataset, a land cover dataset, and an elevation dataset.
12 . The computer system of claim 10 , wherein the high temporal and spatial resolution air temperature forecasts are generated at a real-time (t) or a near real-time, wherein the near real-time comprises a delay in the real-time (t) of less than about 15 minutes.
13 . The computer system of claim 10 , wherein the high temporal and spatial resolution air temperature forecasts are generated at the spatial resolution of 100 meters×100 meters or finer.
14 . A method for generating air temperature forecasts, the method comprising:
providing air temperature data from a multiplicity of sources as a first input to a generative adversarial network;
conditioning the first input to the generative adversarial network on at least one conditioning variable;
training the generative adversarial network with the air temperature data and the conditioning the first input to the generative adversarial network on the at least one conditioning variable;
generating high resolution air temperature maps having a spatial resolution of 100 meters×100 meters or finer as a first output from the generative adversarial network, wherein the high resolution air temperature maps are generated at a higher spatial resolution than the air temperature data from the multiplicity of sources;
providing the high resolution air temperature maps along with real-time air temperature data from a sensor network as a second input to a physics-informed neural network with extended Kalman filter; and
generating high temporal and spatial resolution air temperature forecasts as a second output from the physics-informed neural network with the extended Kalman filter, wherein a loss function of the physics-informed neural network computes loss L as L=L b +L o +L e +L i , wherein L e is a residual loss, L b is a boundary condition loss, L o is an initial condition loss, and L i is a loss term defined by an innovation vector and Kalman gain in the extended Kalman filter.
15 . The method of claim 14 , wherein the multiplicity of sources is selected from a first group consisting of: global climate model forecasts, regional climate model forecasts, urban climate model datasets, and combinations thereof, and wherein the at least one conditioning variable is obtained from a dataset selected from a second group consisting of: a satellite imagery dataset, a land cover dataset, and an elevation dataset.
16 . The method of claim 14 , wherein the first input to the generative adversarial network comprises a first timeseries of the air temperature data from the multiplicity of sources, and wherein the first output from the generative adversarial network comprises a second timeseries of the high resolution air temperature maps.
17 . The method of claim 14 , further comprising:
generating the high resolution air temperature maps using a generator sub-model; and
feeding the high resolution air temperature maps to a discriminator sub-model along with urban climate model datasets having the spatial resolution of 100 meters×100 meters or coarser which are corrected using the real-time air temperature data from the sensor network.
18 . The method of claim 14 , wherein the high temporal and spatial resolution air temperature forecasts are generated at a real-time (t) or a near real-time, wherein the near real-time comprises a delay in the real-time (t) of less than about 15 minutes.
19 . The method of claim 14 , wherein the high temporal and spatial resolution air temperature forecasts are generated at the spatial resolution of 100 meters×100 meters or finer.