Improving performance Metrix in AutoBNN

[Opiyo-Inno]

Hello everyone, i am trying to predict PAR data using the AutoBNN, however it is giving me very big values of MAE, RMSE and R-squared, i request for guidance on the changes to make in the code to make these values be between zero and 1.
here is the code i am running.
import jax
import jax.numpy as jnp
import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
from autobnn import estimators
from autobnn import training_util

Set up seed

seed = jax.random.PRNGKey(0)

Load your PAR data

data = pd.read_excel("D:/Work on PAR/PAR_data.xlsx")
data['date'] = pd.to_datetime(data[['year', 'month', 'day']])
data['value'] = data['PAR (W/m*2)']

Use daily data directly

PAR_daily = data['value'].to_numpy(dtype=np.float32)
num_forecast_steps = 365 * 3 # Forecast the last three years

PAR_training_data = PAR_daily[:-num_forecast_steps]
PAR_dates = data['date'].to_numpy()

fig, ax = plt.subplots(figsize=(16, 4))
ax.plot(PAR_dates[:-num_forecast_steps], PAR_training_data, lw=2)
plt.title("Daily PAR, Location Name")
plt.xlabel("Date")
plt.ylabel("PAR (W/m^2)")

Prepare training and test data

x_test = np.arange(len(PAR_daily), dtype=np.float32)
y_test = PAR_daily
x_train = x_test[:-num_forecast_steps]
y_train = PAR_training_data

x_scale = x_train.max()
x_train = x_train / x_scale
x_test = x_test / x_scale

one_year_period = jnp.array(365.0 / x_scale, dtype=jnp.float32) # daily data

fit_seed, pred_seed, seed = jax.random.split(seed, 3)

Enhanced model with added kernel components and more particles

est = estimators.AutoBnnMapEstimator(
'sum_of_products',
likelihood_model='normal_likelihood_logistic_noise',
seed=jax.random.PRNGKey(0),
periods=(one_year_period,),
num_particles=32, # Increased particles for better approximation
)

Fit the model to the training data

est = est.fit(x_train[..., None], y_train[..., None])

Predict PAR on the test data

preds = est.predict(x_test[..., None])
lo, mid, p90, hi = est.predict_quantiles(x_test[..., None], q=[2.5, 50., 90., 97.5])

_ = training_util.plot_results(PAR_dates,
preds,
dates_test=PAR_dates[-num_forecast_steps:],
y_test=PAR_daily[-num_forecast_steps:],
p2_5=lo,
p50=mid,
p97_5=hi,
dates_train=PAR_dates[:-num_forecast_steps],
y_train=PAR_daily[:-num_forecast_steps],
log_scale=False,
show_particles=False)

ax.set_xlabel("Date")
ax.set_ylabel("PAR (W/m^2)")

Analyze results

absolute_errors = abs(preds - y_test)
mae = absolute_errors.mean()
print("Mean Absolute Error (MAE):", mae)
squared_errors = (preds - y_test) ** 2
mean_squared_error = squared_errors.mean()
rmse = np.sqrt(mean_squared_error)
print("Root Mean Square Error (RMSE):", rmse)

Mean of actual values

y_mean = y_test.mean()

Total Sum of Squares (TSS)

tss = ((y_test - y_mean) ** 2).sum()

Residual Sum of Squares (RSS)

rss = ((y_test - preds) ** 2).sum()

r_squared = 1 - (rss / tss)
print("R-squared:", r_squared)