You can deploy this model and start earning money today!

shahfazal/
nyc-ev-charger-lstm

Name: nyc-ev-charger-lstm
Availability: InStock

time-series-forecasting

mit

Views

No views yet

NYC EV Charger LSTM

LSTM model predicting daily EV charger energy output (kWh) across 8 NYC public charging locations.

Second project in a learning series — progression from a 23-parameter MLP (hello-neural-world) to sequence modeling with LSTMs.

Model versions

Version	File	RMSE	MAE	Params
v0.1	`ev_charger_lstm_v0.1.safetensors`	240.55 kWh	186.11 kWh	4,513
v0.2	`ev_charger_lstm_multilocation.safetensors`	221.31 kWh	137.98 kWh	50,753
v0.3	`ev_charger_lstm_v03.safetensors`	218.98 kWh	136.09 kWh	52,801

See GitHub for architecture details and notebooks for each version.

Latest (v0.3) — Architecture

input_size  = 10      # [kWh_scaled, day_of_week, loc_0...loc_7]
hidden_size = 64
num_layers  = 2
dropout     = 0.2
window_size = 14      # days of lookback
output      = scalar  # next-day kWh

One-hot location encoding (8 binary columns) replaces integer location IDs from v0.2. Day-of-week added as an explicit calendar signal. Dropout added after overfitting was observed without it.

Prophet Baseline (v0.3)

Compared against Facebook Prophet on the same train/test split per location:

Model	RMSE	MAE
Prophet	260.05 kWh	206.22 kWh
LSTM v0.3	218.98 kWh	136.09 kWh

Prophet wins on quiet, stable stations. LSTM generalises better across all 8 — especially on locations with irregular behaviour.

Usage (v0.3)

python

import torch
from safetensors.torch import load_file

class EVChargerLSTM(torch.nn.Module):
    def __init__(self, input_size, hidden_size, num_layers):
        super().__init__()
        self.lstm = torch.nn.LSTM(input_size, hidden_size, num_layers, batch_first=True, dropout=0.2)
        self.fc = torch.nn.Linear(hidden_size, 1)
    def forward(self, x):
        out, _ = self.lstm(x)
        return self.fc(out[:, -1, :]).squeeze(1)

model = EVChargerLSTM(input_size=10, hidden_size=64, num_layers=2)
state_dict = load_file("ev_charger_lstm_v03.safetensors")
model.load_state_dict(state_dict)
model.eval()

# Input: (batch, 14, 10) — [kWh_scaled, day_of_week, loc_0...loc_7]
# Output: (batch,) predicted kWh (scaled) — inverse_transform with per-location scaler

Training data

Source: NYC Open Data — EV charging station sessions
Locations: 8 NYC public charging sites (Hunts Point excluded — 1 row)
Date range: 2021–2026
Split: 80% train / 10% val / 10% test (per location, time-ordered)
Scaling: Per-location MinMaxScaler (fit on train only)

Limitations

No exogenous features (weather, events, holidays)
Trained on NYC data only
Per-location scaler required for inverse transform

shahfazal/nyc-ev-charger-lstm