Pravāh
Blog · November 20, 2026

Building the best time-series model for electricity demand

Jason Chang, Helgi Hilmarsson, Dhruv Suri, and Mohak Mangal

Forecasting electricity demand is a critical task undertaken by grid operators around the world. With Banyan, we set out to build the best time-series model for this task. In a year of testing against 10 other forecasting approaches, including frontier time-series foundation models, Banyan had the lowest error in all 7 US ISOs.

Across the seven regional systems, its average error was nearly 25% lower than the closest competitor.

Two-day forecast error across seven ISO systems

Banyan1.44%
The three lowest scores, averaged equally across the seven systems. Shorter bars indicate lower weighted absolute percentage error (WAPE). The appendix includes all 11 approaches.

The importance of forecasting electricity

Grid operators must decide how much electricity to have ready before homes and businesses need it. They use demand forecasts to schedule generators, arrange power purchases, and prepare for peaks. If demand comes in higher than expected, they may have to find extra supply at short notice. If it comes in lower, they may have paid to start generators they did not need.

Inaccurate forecasting can have serious consequences. During Winter Storm Elliott in December 2022, the Tennessee Valley Authority ordered rolling outages on December 23 and 24 as demand exceeded forecasts and generating units failed. In Nashville, neighborhoods lost power for about 10 minutes every 90 minutes to two hours during the rotations. Around 72,000 people in Nashville Electric Service’s territory were without power at the storm’s peak. The subsequent FERC and NERC inquiry found that many operators across the affected region had underestimated peak demand.

Ice-covered trees, fallen branches, and downed utility lines across a Tennessee street after a February 2022 winter storm.

Downed utility lines and ice-covered trees after a February 2022 winter storm in Tennessee. Photo: Brad Vest/Getty Images. Source: The New York Times.

Testing a full year of electricity demand

We tested Banyan against 10 other forecasting approaches, including frontier time-series foundation models, from October 2025 through September 2026. Every six hours, each predicted demand for each of the next 48 hours. The test covered the seven U.S. regional grid systems and 80 zones within them. We scored the predictions against each operator’s published demand, collected through Grid Status.

We ran the test as though each forecast were being made at the time. Models received the weather forecasts and demand readings that would have been available then. We kept the test year out of model training and scored every approach on the same forecast hours.

Across all 87 areas, Banyan’s average error was 2.47%, compared with 2.73% for fine-tuned Timestral and 3.21% for fine-tuned Chronos-2. TiDE, an MLP-based model and the strongest model trained from scratch for the comparison, scored 3.25%. Banyan beat TiDE in every area. We measured each area’s error relative to its total demand, then averaged those scores with equal weight for each area.

Forecast error by model

Average WAPE across the 7 regional systems. Lower is better.

Banyan
1.44%Lowest average error
Timestral · fine-tuned
1.91%32.6% more error than Banyan
MLP-based model (TiDE)
2.16%50.0% more error than Banyan
Chronos-2 · fine-tuned
2.27%57.6% more error than Banyan
TFT
2.62%81.9% more error than Banyan
Timestral · zero-shot
2.63%82.6% more error than Banyan
TimesFM 3 · zero-shot
2.69%86.8% more error than Banyan
TimesFM 3 · fine-tuned
2.76%91.7% more error than Banyan
Chronos-2 · zero-shot
3.14%118.1% more error than Banyan
DAG
3.30%129.2% more error than Banyan
LightGBM
3.65%153.5% more error than Banyan
0%3%6%

Figure 2. All 11 models. Select the seven regional systems, 80 zones, or all 87 areas. WAPE measures absolute forecast error relative to demand. Each area receives equal weight.

Banyan leads across regions, seasons, and forecast horizons

Banyan had the lowest average error in every month of the test year. Across all 87 areas, it also led at every forecast horizon from two to 48 hours ahead. At 48 hours, its error was 2.73%, compared with 3.11% for fine-tuned Timestral. For the seven regional systems alone, Banyan led from three to 48 hours ahead.

Forecast error on each regional system

BanyanTimestral, fine-tunedMLP-based model (TiDE)
0%1%2%3%CAISO · Banyan · 2.17% WAPE2.17CAISO · Timestral, fine-tuned · 2.32% WAPE2.32CAISO · TiDE · 2.72% WAPE2.72CAISOERCOT · Banyan · 1.36% WAPE1.36ERCOT · Timestral, fine-tuned · 1.50% WAPE1.50ERCOT · TiDE · 1.81% WAPE1.81ERCOTISO-NE · Banyan · 1.93% WAPE1.93ISO-NE · Timestral, fine-tuned · 2.71% WAPE2.71ISO-NE · TiDE · 2.82% WAPE2.82ISO-NEMISO · Banyan · 0.93% WAPE0.93MISO · Timestral, fine-tuned · 1.70% WAPE1.70MISO · TiDE · 1.90% WAPE1.90MISONYISO · Banyan · 1.45% WAPE1.45NYISO · Timestral, fine-tuned · 2.09% WAPE2.09NYISO · TiDE · 2.07% WAPE2.07NYISOPJM · Banyan · 1.06% WAPE1.06PJM · Timestral, fine-tuned · 1.38% WAPE1.38PJM · TiDE · 1.93% WAPE1.93PJMSPP · Banyan · 1.17% WAPE1.17SPP · Timestral, fine-tuned · 1.67% WAPE1.67SPP · TiDE · 1.87% WAPE1.87SPP

Figure 3. System-wide error in each region. Banyan, fine-tuned Timestral, and TiDE are shown. Banyan ranks first among all 11 models on each system.

Forecasting during extreme weather periods

A forecast’s performance during extreme weather matters because demand can rise sharply while supply is under pressure. Extreme heat and cold can change demand quickly as heating and cooling systems run harder. Testing these periods shows whether a model remains accurate when operators have less room for error. In PJM’s January 20–27, 2026 cold spell, Banyan’s error was 1.15%, compared with 1.46% for fine-tuned Timestral and 2.43% for TiDE. The chart shows actual demand and predictions made roughly a day earlier.

Actual loadBanyanBest fine-tuned foundation model (Timestral)MLP-based model (TiDE)
coldTue 20 JancoldWed 21 JanThu 22 JancoldFri 23 JancoldSat 24 Jancold + stormSun 25 JancoldMon 26 JancoldTue 27 Jan100120140GW−100+10error %

January 20–27, 2026. The upper panel shows demand in gigawatts. The lower panel shows forecast error. Shading identifies days flagged for cold weather.

How we built Banyan

Banyan is a transformer-based model built for electricity demand forecasting. We train it across multiple grid regions, then fine-tune it for individual areas. Its training also includes auxiliary pre-training tasks.

Electricity demand depends on where people live and what the weather is doing there. A heat wave over a city can drive much more air-conditioning use than the same heat over open land. Averaging the weather across a large region can hide that difference.

Side-by-side pixel-art maps. Heat falls over the populated city on the left and open land on the right, despite the same regional average temperature.

Heat over a city and heat over open land can produce different electricity demand even when the regional average temperature is the same. Conceptual illustration.

Banyan uses weather maps to retain those local differences. It combines them with past electricity demand, population, terrain, and installed solar and wind capacity to predict demand hour by hour. The results above show that this approach can outperform general-purpose time-series models, including versions fine-tuned on electricity data, on this benchmark.

Important research contributions

  1. Spatial inputs for electricity demand: Banyan combines weather maps with population, terrain, and installed solar and wind capacity. This gives the model information about where weather occurs and the electricity system it affects.
  2. A year-long evaluation under realistic operating conditions: We tested forecasts across seven regional systems and 80 zones using the information available when each forecast would have been issued. The evaluation includes seasonal changes, cold spells, and peak-demand periods.
  3. A direct comparison with adapted foundation models: We evaluated Timestral, Chronos-2, and TimesFM 3 both as released and after fine-tuning on electricity data. Fine-tuning reduced Timestral’s average error from 3.84% to 2.73%, making it Banyan’s closest competitor. Banyan led across all seven regional systems and in 77 of 87 areas even after competing models had learned from electricity data.

Appendix: How we trained the models and measured error

We used January 2019 through September 2024 for training and the following year for validation. Banyan was fitted through September 2025. The fine-tuned foundation models trained through September 2024, with the following year used to choose a checkpoint. After validation, the models trained from scratch were refitted through September 2025. Every model stayed fixed during the test year.

The foundation models, TFT, and DAG received load history, area-average weather, and calendar inputs. TiDE and LightGBM used weather from 184 stations. Banyan used weather maps and geographic inputs. This benchmark compares the complete approaches. Establishing how much of Banyan’s lead comes from each input or modeling choice will require additional experiments.

We issued forecasts at 00, 06, 12, and 18 UTC for the next 48 hours. We used the latest available ECMWF AIFS weather forecast and respected each area’s demand-reporting delay, which ranged from one to 61 hours. Each approach was scored on 6,052,976 matched forecast hours.

We report weighted absolute percentage error (WAPE). For each area, we divided the sum of absolute forecast errors by total observed demand. We then averaged the area scores equally. The seven-system score gives each regional system equal weight. The 87-area score gives a zone the same weight as a regional system.

Full results by model and grid area

Explore the benchmark tables

All 11 approaches · average WAPE (%)

ModelAll 87 areas7 ISO systems80 zonesError vs BanyanBest in (areas)
Banyan2.47 ●1.44 ●2.57 ●—77
Timestral 209M, fine-tunedMistral foundation model2.731.912.81+10 %9
Chronos-2 120M, fine-tunedAmazon foundation model3.212.273.30+30 %1
TiDEGoogle, 20233.252.163.35+31 %0
TFTGoogle, 20193.652.623.74+47 %0
TimesFM 3 331M, fine-tunedGoogle foundation model3.762.763.85+52 %0
TimesFM 3 331M, zero-shotGoogle foundation model3.792.693.89+53 %0
Timestral 209M, zero-shotMistral foundation model3.842.633.94+55 %0
LightGBMgradient boosting4.383.654.44+77 %0
Chronos-2 120M, zero-shotAmazon foundation model4.593.144.72+86 %0
DAGICML 20264.623.304.73+87 %0

All 87 areas · WAPE (%) by model

FT means fine-tuned on our data. ZS means zero-shot. A dot marks the lowest error using the underlying scores before rounding.

CAISO

AreaBanyanTimestral FTChronos-2 FTTimesFM 3 FT
CAISO system2.17 ●2.322.523.03
PG&E2.28 ●2.542.783.39
San Diego Gas & Electric5.295.05 ●5.215.90
Southern California Edison3.303.22 ●3.453.95
Valley Electric4.02 ●4.535.095.76
AreaTimestral ZSChronos-2 ZSTimesFM 3 ZSMLP-based model (TiDE)
CAISO system2.762.912.802.72
PG&E3.033.253.072.81
San Diego Gas & Electric5.505.645.535.82
Southern California Edison3.623.893.733.70
Valley Electric5.395.945.435.21
AreaTFTDAGLightGBM
CAISO system3.553.174.18
PG&E2.963.283.45
San Diego Gas & Electric6.956.017.62
Southern California Edison4.514.195.67
Valley Electric6.566.926.12

ERCOT

AreaBanyanTimestral FTChronos-2 FTTimesFM 3 FT
ERCOT system1.36 ●1.501.791.76
Coast2.00 ●2.112.432.51
East3.36 ●3.574.034.10
Far West2.992.752.75 ●3.13
Houston load zone2.00 ●2.112.432.51
North5.324.72 ●4.865.16
North Central2.13 ●2.302.822.80
North load zone1.98 ●2.112.612.61
South2.28 ●2.332.602.65
South Central1.96 ●2.142.572.69
South load zone1.61 ●1.732.062.16
West6.345.51 ●5.966.33
West load zone2.402.20 ●2.232.64
AreaTimestral ZSChronos-2 ZSTimesFM 3 ZSMLP-based model (TiDE)
ERCOT system2.012.691.981.81
Coast2.683.342.652.37
East4.445.484.444.16
Far West2.792.872.833.45
Houston load zone2.683.332.652.37
North4.985.165.055.66
North Central3.204.283.112.53
North load zone3.004.052.932.32
South2.813.362.772.56
South Central2.963.962.932.40
South load zone2.413.262.351.85
West6.086.566.107.56
West load zone2.292.412.342.77
AreaTFTDAGLightGBM
ERCOT system1.763.223.05
Coast2.523.123.74
East4.786.255.40
Far West3.043.123.35
Houston load zone2.523.113.74
North5.956.366.24
North Central2.915.315.63
North load zone2.624.895.17
South2.893.444.43
South Central2.964.294.76
South load zone2.423.334.17
West6.759.527.96
West load zone2.552.812.76

ISO-NE

AreaBanyanTimestral FTChronos-2 FTTimesFM 3 FT
ISO-NE system1.93 ●2.713.254.28
Connecticut2.41 ●2.793.574.47
Maine2.80 ●3.824.506.06
New Hampshire2.09 ●2.473.144.20
Northeast Mass.1.81 ●2.262.923.67
Rhode Island3.63 ●3.794.405.18
Southeast Mass.2.92 ●3.464.225.24
Vermont4.19 ●4.935.717.77
Western/Central Mass.2.45 ●2.983.654.69
AreaTimestral ZSChronos-2 ZSTimesFM 3 ZSMLP-based model (TiDE)
ISO-NE system4.124.604.122.82
Connecticut5.146.294.923.60
Maine5.956.065.824.29
New Hampshire4.275.364.383.10
Northeast Mass.4.064.903.733.04
Rhode Island5.666.855.674.95
Southeast Mass.5.566.595.264.21
Vermont6.987.307.105.37
Western/Central Mass.5.256.145.233.46
AreaTFTDAGLightGBM
ISO-NE system3.604.864.25
Connecticut3.325.604.29
Maine5.406.385.35
New Hampshire2.984.443.73
Northeast Mass.3.204.463.33
Rhode Island5.586.705.06
Southeast Mass.4.625.984.75
Vermont6.547.456.50
Western/Central Mass.4.245.073.86

MISO

AreaBanyanTimestral FTChronos-2 FTTimesFM 3 FT
MISO system0.93 ●1.701.992.44
Zone 11.75 ●2.482.943.55
Zone 42.20 ●2.503.003.79
Zone 61.91 ●2.072.593.39
Zones 2 and 71.33 ●2.833.444.28
Zones 3 and 51.54 ●2.432.893.41
Zones 8, 9, and 101.66 ●2.392.742.86
AreaTimestral ZSChronos-2 ZSTimesFM 3 ZSMLP-based model (TiDE)
MISO system2.192.602.351.90
Zone 13.784.083.763.13
Zone 44.565.234.213.29
Zone 63.464.623.582.75
Zones 2 and 74.004.344.232.76
Zones 3 and 53.834.883.842.80
Zones 8, 9, and 103.164.593.253.43
AreaTFTDAGLightGBM
MISO system2.172.532.62
Zone 13.133.564.42
Zone 43.774.965.80
Zone 63.373.973.46
Zones 2 and 73.564.053.01
Zones 3 and 53.214.204.55
Zones 8, 9, and 103.715.114.87

NYISO

AreaBanyanTimestral FTChronos-2 FTTimesFM 3 FT
NYISO system1.45 ●2.092.503.27
Capital2.97 ●3.303.794.78
Central2.90 ●3.073.504.43
Dunwoodie2.35 ●2.593.093.73
Genesee2.69 ●2.963.454.56
Hudson Valley2.76 ●3.233.804.64
Long Island2.50 ●3.133.924.86
Millwood3.96 ●4.094.695.36
Mohawk Valley4.02 ●4.174.646.06
New York City1.42 ●1.722.172.64
North3.433.31 ●3.423.98
West2.33 ●2.342.713.46
AreaTimestral ZSChronos-2 ZSTimesFM 3 ZSMLP-based model (TiDE)
NYISO system3.013.453.102.07
Capital4.584.864.563.81
Central4.264.514.243.88
Dunwoodie3.864.373.603.39
Genesee4.244.604.243.47
Hudson Valley4.625.324.653.53
Long Island4.314.744.313.28
Millwood5.416.215.234.97
Mohawk Valley5.505.685.404.99
New York City2.563.252.571.87
North3.673.593.553.91
West3.023.303.022.95
AreaTFTDAGLightGBM
NYISO system2.803.573.22
Capital4.785.535.13
Central4.255.335.07
Dunwoodie3.884.513.87
Genesee3.594.964.40
Hudson Valley4.265.375.14
Long Island3.935.584.36
Millwood5.376.196.24
Mohawk Valley5.286.466.55
New York City2.223.122.63
North4.254.394.44
West3.223.794.26

PJM

AreaBanyanTimestral FTChronos-2 FTTimesFM 3 FT
PJM system1.06 ●1.381.822.29
AEP1.27 ●1.441.822.27
ATSI1.64 ●1.782.242.73
Allegheny Power1.60 ●1.862.372.77
Atlantic Electric3.02 ●3.443.984.82
BGE2.34 ●2.543.153.60
ComEd2.00 ●2.152.603.25
Dayton1.93 ●2.122.693.40
Delmarva2.35 ●2.843.514.00
Dominion1.35 ●1.642.032.19
Duke Ohio/Kentucky2.06 ●2.202.773.25
Duquesne2.10 ●2.222.522.98
East Kentucky4.16 ●4.514.905.39
JCP&L2.66 ●3.013.614.32
Met-Ed1.81 ●2.082.703.27
Mid-Atlantic region1.57 ●1.872.462.98
PECO1.95 ●2.112.633.06
PPL1.70 ●1.832.402.84
PSE&G2.13 ●2.392.963.68
Penelec1.58 ●1.722.192.86
Pepco2.29 ●2.583.113.71
Rockland Electric2.42 ●2.663.183.86
Western region1.13 ●1.331.782.30
AreaTimestral ZSChronos-2 ZSTimesFM 3 ZSMLP-based model (TiDE)
PJM system2.112.952.201.93
AEP2.182.802.281.72
ATSI2.643.242.762.24
Allegheny Power2.793.612.852.03
Atlantic Electric4.565.664.573.85
BGE3.834.963.622.96
ComEd3.223.973.262.67
Dayton3.224.043.362.72
Delmarva3.995.263.802.85
Dominion2.383.372.271.84
Duke Ohio/Kentucky3.384.273.312.64
Duquesne3.073.873.092.76
East Kentucky5.456.415.314.84
JCP&L4.425.274.413.16
Met-Ed3.163.973.322.46
Mid-Atlantic region2.963.942.992.17
PECO3.144.103.162.46
PPL2.813.652.942.25
PSE&G3.734.383.692.65
Penelec2.423.032.692.10
Pepco3.984.993.642.89
Rockland Electric3.704.533.852.98
Western region2.142.862.291.85
AreaTFTDAGLightGBM
PJM system2.142.843.65
AEP1.893.465.25
ATSI2.283.422.80
Allegheny Power2.483.372.83
Atlantic Electric5.175.804.45
BGE3.544.443.69
ComEd2.994.023.49
Dayton2.973.803.22
Delmarva3.895.033.72
Dominion2.253.284.42
Duke Ohio/Kentucky3.063.993.03
Duquesne2.793.543.21
East Kentucky5.045.985.72
JCP&L3.755.014.17
Met-Ed3.213.732.89
Mid-Atlantic region2.793.772.76
PECO2.823.742.96
PPL2.903.632.81
PSE&G3.134.403.58
Penelec2.293.012.51
Pepco3.764.473.69
Rockland Electric3.704.353.79
Western region1.912.973.59

SPP

AreaBanyanTimestral FTChronos-2 FTTimesFM 3 FT
SPP system1.17 ●1.672.032.26
AEP West1.97 ●2.222.783.02
City of Springfield2.82 ●3.073.724.15
Empire District2.92 ●3.174.024.44
Evergy Kansas Central1.95 ●2.182.773.15
Evergy Metro2.79 ●2.993.534.10
Evergy Missouri West3.08 ●3.734.594.92
Grand River Dam Auth.1.851.77 ●1.992.06
Independence P&L3.82 ●3.974.615.01
Kansas City BPU3.04 ●3.123.744.51
Lincoln Electric3.573.49 ●4.204.54
Nebraska PPD4.76 ●5.416.156.84
OG&E2.46 ●2.583.063.24
Omaha PPD3.402.98 ●3.523.88
Southwestern Public Service1.65 ●1.962.182.34
Sunflower Electric2.74 ●2.983.333.55
WAPA Upper Great Plains1.60 ●1.862.052.27
Western Farmers2.44 ●2.583.003.27
AreaTimestral ZSChronos-2 ZSTimesFM 3 ZSMLP-based model (TiDE)
SPP system2.242.772.311.87
AEP West3.544.983.432.70
City of Springfield4.936.034.624.01
Empire District5.086.674.854.55
Evergy Kansas Central3.664.723.552.84
Evergy Metro4.825.934.323.97
Evergy Missouri West5.907.325.264.31
Grand River Dam Auth.2.352.762.282.59
Independence P&L5.417.485.304.73
Kansas City BPU4.675.964.744.16
Lincoln Electric5.267.024.994.66
Nebraska PPD6.937.996.735.50
OG&E3.844.553.552.91
Omaha PPD4.235.444.083.67
Southwestern Public Service2.362.592.302.18
Sunflower Electric3.754.053.663.47
WAPA Upper Great Plains2.572.822.462.54
Western Farmers3.254.633.333.19
AreaTFTDAGLightGBM
SPP system2.292.954.59
AEP West3.353.923.61
City of Springfield4.185.564.78
Empire District5.735.865.93
Evergy Kansas Central2.974.082.91
Evergy Metro4.816.045.26
Evergy Missouri West4.895.895.02
Grand River Dam Auth.2.206.314.34
Independence P&L4.766.244.95
Kansas City BPU4.565.404.93
Lincoln Electric4.155.567.83
Nebraska PPD6.187.926.53
OG&E3.104.866.39
Omaha PPD3.814.695.42
Southwestern Public Service2.062.402.57
Sunflower Electric3.363.744.74
WAPA Upper Great Plains2.813.512.87
Western Farmers3.084.353.51

Model references

  • Timestral — Mistral’s time-series foundation models.
  • Chronos-2 — Amazon’s Chronos forecasting models.
  • TimesFM 3 — Google Research’s TimesFM models.
  • TiDE — Long-term Forecasting with TiDE: Time-series Dense Encoder.
  • TFT — Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting.
  • LightGBM — Microsoft Research’s LightGBM project.
  • DAG — A Dual Correlation Network for Time Series Forecasting with Exogenous Variables.