Reproducing the "discharge" and "full" model

[fingertap]

Hi, thanks for this very good and organized codebase.

Following your code, the "variance" model is easy to reproduce. However, I have trouble reproducing the performance of the "discharge" and "full" model. Do you know what detail I may be missing?

My results are:

(103.5718828346258, 138.2982760624032, 195.8641688228285)    # var
(62.6994702955215, 181.47910490241762, 194.8456325157642)    # discharge
(62.61943481775296, 141.89261700359316, 170.8711847567794)    # full

Here is my code for feature generation:

def var_feature(batch, to_skip=None):
    feats = []
    to_skip = to_skip or []
    if isinstance(to_skip, int):
        to_skip = [to_skip]
    for indx, cell in enumerate(batch):
        if indx in to_skip:
            continue
        x = cell['cycles']['99']['Qdlin'] - cell['cycles']['9']['Qdlin']
        feats.append([np.log10(np.var(x))])
    return np.array(feats)

def discharge_feature(batch, to_skip=None, freq=10):
    feats = []
    to_skip = to_skip or []
    if isinstance(to_skip, int):
        to_skip = [to_skip]
    for indx, cell in enumerate(batch):
        if indx in to_skip:
            continue
        cell_feat = []
        x = cell['cycles']['99']['Qdlin'] - cell['cycles']['9']['Qdlin']
        cell_feat.append(np.var(x[::freq]))
        cell_feat.append(np.abs(np.min(x[::freq])))
        cell_feat.append(np.abs(skew(x[::freq])))
        cell_feat.append(np.abs(kurtosis(x[::freq])))
        cell_feat.append(cell['summary']['QD'][1])
        cell_feat.append(cell['summary']['QD'][1:100].max() - cell['summary']['QD'][1])
        feats.append(np.log10(cell_feat))
    return np.array(feats)

def full_feature(batch, to_skip=None):
    feats = []
    to_skip = to_skip or []
    if isinstance(to_skip, int):
        to_skip = [to_skip]
    for indx, cell in enumerate(batch):
        if indx in to_skip:
            continue
        cell_feat = []
        x = cell['cycles']['99']['Qdlin'] - cell['cycles']['9']['Qdlin']
        cell_feat.append(np.log10(np.var(x)))
        cell_feat.append(np.log10(np.abs(np.min(x))))
        m = LinearRegression().fit(
            np.ones((99, 1)),
            cell['summary']['QD'][1:100])
        cell_feat.append(np.abs(float(m.coef_)))
        cell_feat.append(np.abs(m.intercept_))
        cell_feat.append(np.log10(cell['summary']['QD'][1]))
        cell_feat.append(np.log10(cell['summary']['chargetime'][:5].mean()))
        cell_feat.append(np.log10(cell['summary']['Tavg'][1:100].sum()))
        cell_feat.append(np.log10(cell['summary']['IR'][1:100].min() + 1e-8))
        cell_feat.append(np.log10(abs(cell['summary']['IR'][100] - cell['summary']['IR'][1])))
        feats.append(cell_feat)

    return np.array(feats)

I just use a linear regression after the standard scaler:

def train(x_train, x_test1, x_test2):
    scaler = preprocessing.StandardScaler().fit(x_train)
    x_train = scaler.transform(x_train)
    x_test1 = scaler.transform(x_test1)
    x_test2 = scaler.transform(x_test2)

    # Define and fit linear regression via enet
    # l1_ratios = [0.1, 0.5, 0.7, 0.9, 0.95, 0.99, 1]
    # enet = ElasticNetCV(l1_ratio=l1_ratios, cv=5, random_state=0)
    enet = LinearRegression()
    enet.fit(x_train, y_train)

    # Predict on test sets
    y_train_pred = enet.predict(x_train)
    y_test1_pred = enet.predict(x_test1)
    y_test2_pred = enet.predict(x_test2)

    # Evaluate error
    return get_RMSE_for_all_datasets(y_train_pred, y_test1_pred, y_test2_pred)

[petermattia]

Hi, sorry for the delayed response. Just to confirm, you're saying your run of the notebook gave you different results than what is posted here?

[fingertap]

Actually, I can reproduce all your results in the paper “statistical learning for accurate and interpretable battery lifetime prediction”, including the variance model. However, I cannot reproduce the discharge and full model from the Nature Energy paper. I cannot find it in this repo either.

[petermattia]

Unfortunately, you need the license from Prof. Braatz for access to this code (see the readme). Doubly unfortunately, he rarely responds to requests for this license. Sorry about that.